Vendor OpenClaw source as Adolf fork baseline

Adolf is a fork/vendored clone of github.com/openclaw/openclaw (v2026.6.11),
free to diverge. Tree copied sans upstream .git; upstream remote added for
future syncs. Node pinned to 24 (.nvmrc); engines already require >=22.19.
Preserves docs/ARCHITECTURE.md.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LeqyaxJF2nbRXJtae2kNB2

packages/memory-host-sdk/src/engine-embeddings.ts

@@ -0,0 +1,74 @@
+// Real workspace contract for memory embedding providers and batch helpers.
+
+export {
+  getMemoryEmbeddingProvider,
+  listRegisteredMemoryEmbeddingProviders,
+  listMemoryEmbeddingProviders,
+  listRegisteredMemoryEmbeddingProviderAdapters,
+} from "./host/openclaw-runtime-memory.js";
+export type {
+  MemoryEmbeddingBatchChunk,
+  MemoryEmbeddingBatchOptions,
+  MemoryEmbeddingProvider,
+  MemoryEmbeddingProviderAdapter,
+  MemoryEmbeddingProviderCallOptions,
+  MemoryEmbeddingProviderCreateOptions,
+  MemoryEmbeddingProviderCreateResult,
+  MemoryEmbeddingProviderRuntime,
+} from "./host/openclaw-runtime-memory.js";
+export { createLocalEmbeddingProvider, DEFAULT_LOCAL_MODEL } from "./host/embeddings.js";
+export { extractBatchErrorMessage, formatUnavailableBatchError } from "./host/batch-error-utils.js";
+export { postJsonWithRetry } from "./host/batch-http.js";
+export { applyEmbeddingBatchOutputLine } from "./host/batch-output.js";
+export {
+  EMBEDDING_BATCH_ENDPOINT,
+  type EmbeddingBatchStatus,
+  type ProviderBatchOutputLine,
+} from "./host/batch-provider-common.js";
+export {
+  buildEmbeddingBatchGroupOptions,
+  runEmbeddingBatchGroups,
+  type EmbeddingBatchExecutionParams,
+} from "./host/batch-runner.js";
+export {
+  resolveBatchCompletionFromStatus,
+  resolveCompletedBatchResult,
+  throwIfBatchTerminalFailure,
+  type BatchCompletionResult,
+} from "./host/batch-status.js";
+export { uploadBatchJsonlFile } from "./host/batch-upload.js";
+export {
+  buildBatchHeaders,
+  normalizeBatchBaseUrl,
+  type BatchHttpClientConfig,
+} from "./host/batch-utils.js";
+export { enforceEmbeddingMaxInputTokens } from "./host/embedding-chunk-limits.js";
+export {
+  isMissingEmbeddingApiKeyError,
+  mapBatchEmbeddingsByIndex,
+  sanitizeEmbeddingCacheHeaders,
+} from "./host/embedding-provider-adapter-utils.js";
+export { sanitizeAndNormalizeEmbedding } from "./host/embedding-vectors.js";
+export { debugEmbeddingsLog } from "./host/embeddings-debug.js";
+export { normalizeEmbeddingModelWithPrefixes } from "./host/embeddings-model-normalize.js";
+export {
+  resolveRemoteEmbeddingBearerClient,
+  type RemoteEmbeddingProviderId,
+} from "./host/embeddings-remote-client.js";
+export {
+  createRemoteEmbeddingProvider,
+  resolveRemoteEmbeddingClient,
+  type RemoteEmbeddingClient,
+} from "./host/embeddings-remote-provider.js";
+export { fetchRemoteEmbeddingVectors } from "./host/embeddings-remote-fetch.js";
+export {
+  estimateStructuredEmbeddingInputBytes,
+  estimateUtf8Bytes,
+} from "./host/embedding-input-limits.js";
+export { hasNonTextEmbeddingParts, type EmbeddingInput } from "./host/embedding-inputs.js";
+export { buildRemoteBaseUrlPolicy, withRemoteHttpResponse } from "./host/remote-http.js";
+export {
+  buildCaseInsensitiveExtensionGlob,
+  classifyMemoryMultimodalPath,
+  getMemoryMultimodalExtensions,
+} from "./host/multimodal.js";

packages/memory-host-sdk/src/engine-foundation.ts

@@ -0,0 +1,49 @@
+// Real workspace contract for memory engine foundation concerns.
+
+export {
+  resolveAgentContextLimits,
+  resolveAgentDir,
+  resolveAgentWorkspaceDir,
+  resolveDefaultAgentId,
+  resolveSessionAgentId,
+} from "./host/openclaw-runtime-agent.js";
+export {
+  resolveMemorySearchConfig,
+  resolveMemorySearchSyncConfig,
+  type ResolvedMemorySearchConfig,
+  type ResolvedMemorySearchSyncConfig,
+} from "./host/openclaw-runtime-agent.js";
+export { parseDurationMs } from "./host/openclaw-runtime-config.js";
+export { loadConfig } from "./host/openclaw-runtime-config.js";
+export { resolveStateDir } from "./host/openclaw-runtime-config.js";
+export { resolveSessionTranscriptsDirForAgent } from "./host/openclaw-runtime-config.js";
+export {
+  hasConfiguredSecretInput,
+  normalizeResolvedSecretInputString,
+} from "./host/openclaw-runtime-config.js";
+export { root } from "./host/openclaw-runtime-io.js";
+export { isPathInside } from "./host/fs-utils.js";
+export { createSubsystemLogger } from "./host/openclaw-runtime-io.js";
+export { detectMime } from "./host/openclaw-runtime-io.js";
+export { resolveGlobalSingleton } from "./host/openclaw-runtime-io.js";
+export { onSessionTranscriptUpdate } from "./host/openclaw-runtime-session.js";
+export { splitShellArgs } from "./host/openclaw-runtime-io.js";
+export { runTasksWithConcurrency } from "./host/openclaw-runtime-io.js";
+export {
+  shortenHomeInString,
+  shortenHomePath,
+  resolveUserPath,
+  truncateUtf16Safe,
+} from "./host/openclaw-runtime-io.js";
+export type { OpenClawConfig } from "./host/openclaw-runtime-config.js";
+export type { SessionSendPolicyConfig } from "./host/openclaw-runtime-config.js";
+export type { SecretInput } from "./host/openclaw-runtime-config.js";
+export type {
+  MemoryBackend,
+  MemoryCitationsMode,
+  MemoryQmdConfig,
+  MemoryQmdIndexPath,
+  MemoryQmdMcporterConfig,
+  MemoryQmdSearchMode,
+} from "./host/openclaw-runtime-config.js";
+export type { MemorySearchConfig } from "./host/openclaw-runtime-config.js";

packages/memory-host-sdk/src/engine-qmd.ts

@@ -0,0 +1,41 @@
+// Real workspace contract for QMD/session/query helpers used by the memory engine.
+
+export { extractKeywords, isQueryStopWordToken } from "./host/query-expansion.js";
+export {
+  buildSessionEntry,
+  listSessionFilesForAgent,
+  listSessionTranscriptCorpusEntriesForAgent,
+  loadDreamingNarrativeTranscriptPathSetForAgent,
+  loadSessionTranscriptClassificationForAgent,
+  normalizeSessionTranscriptPathForComparison,
+  parseCanonicalSessionSyncTargetFromPath,
+  resolveSessionIdentityForTranscriptFile,
+  resolveSessionFileForSyncTarget,
+  sessionPathForFile,
+  type BuildSessionEntryOptions,
+  type ResolvedMemorySessionSyncTarget,
+  type ResolvedSessionTranscriptIdentity,
+  type SessionFileEntry,
+  type SessionTranscriptClassification,
+  type SessionTranscriptCorpusEntry,
+} from "./host/session-files.js";
+export {
+  isSessionArchiveArtifactName,
+  isUsageCountedSessionTranscriptFileName,
+  parseUsageCountedSessionIdFromFileName,
+} from "./host/openclaw-runtime-session.js";
+export { parseQmdQueryJson, type QmdQueryResult } from "./host/qmd-query-parser.js";
+export {
+  deriveQmdScopeChannel,
+  deriveQmdScopeChatType,
+  isQmdScopeAllowed,
+} from "./host/qmd-scope.js";
+export {
+  checkQmdBinaryAvailability,
+  resolveCliSpawnInvocation,
+  resolveQmdBinaryUnavailableReason,
+  runCliCommand,
+  type QmdBinaryAvailability,
+  type QmdBinaryUnavailable,
+  type QmdBinaryUnavailableReason,
+} from "./host/qmd-process.js";

packages/memory-host-sdk/src/engine-storage.ts

@@ -0,0 +1,60 @@
+// Real workspace contract for memory engine storage/index helpers.
+
+export {
+  buildFileEntry,
+  buildMultimodalChunkForIndexing,
+  chunkMarkdown,
+  cosineSimilarity,
+  ensureDir,
+  hashText,
+  listMemoryFiles,
+  normalizeExtraMemoryPaths,
+  parseEmbedding,
+  remapChunkLines,
+  runWithConcurrency,
+  type MemoryChunk,
+  type MemoryFileEntry,
+} from "./host/internal.js";
+export { readMemoryFile } from "./host/read-file.js";
+export { isTransientMemoryReadError, retryTransientMemoryRead } from "./host/read-retry.js";
+export {
+  buildMemoryReadResult,
+  buildMemoryReadResultFromSlice,
+  DEFAULT_MEMORY_READ_LINES,
+  DEFAULT_MEMORY_READ_MAX_CHARS,
+  type MemoryReadResult,
+} from "./host/read-file-shared.js";
+export { resolveMemoryBackendConfig } from "./host/backend-config.js";
+export type {
+  ResolvedMemoryBackendConfig,
+  ResolvedQmdConfig,
+  ResolvedQmdMcporterConfig,
+} from "./host/backend-config.js";
+export type {
+  MemoryEmbeddingProbeResult,
+  MemoryProviderStatus,
+  MemorySearchManager,
+  MemorySearchRuntimeDebug,
+  MemorySearchResult,
+  MemorySessionSyncTarget,
+  MemorySource,
+  MemorySyncParams,
+  MemorySyncProgressUpdate,
+} from "./host/types.js";
+export {
+  ensureMemoryIndexSchema,
+  MEMORY_EMBEDDING_CACHE_TABLE,
+  MEMORY_INDEX_CHUNKS_TABLE,
+  MEMORY_INDEX_FTS_TABLE,
+  MEMORY_INDEX_META_TABLE,
+  MEMORY_INDEX_SOURCES_TABLE,
+  MEMORY_INDEX_STATE_TABLE,
+  MEMORY_INDEX_VECTOR_TABLE,
+} from "./host/memory-schema.js";
+export { loadSqliteVecExtension } from "./host/sqlite-vec.js";
+export {
+  closeMemorySqliteWalMaintenance,
+  configureMemorySqliteWalMaintenance,
+  requireNodeSqlite,
+} from "./host/sqlite.js";
+export { isFileMissingError, statRegularFile } from "./host/fs-utils.js";

packages/memory-host-sdk/src/engine.ts

@@ -0,0 +1,7 @@
+// Aggregate workspace contract for the memory engine surface.
+// Keep focused subpaths preferred for new code.
+
+export * from "./engine-foundation.js";
+export * from "./engine-storage.js";
+export * from "./engine-embeddings.js";
+export * from "./engine-qmd.js";

packages/memory-host-sdk/src/host/backend-config.test.ts

@@ -0,0 +1,772 @@
+// Memory Host SDK tests cover backend config behavior.
+import syncFs from "node:fs";
+import type { Dirent } from "node:fs";
+import fs from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { afterAll, beforeAll, describe, expect, it, vi } from "vitest";
+import { resolveMemoryBackendConfig } from "./backend-config.js";
+import type { OpenClawConfig } from "./config-utils.js";
+
+type ResolvedMemoryBackendConfig = ReturnType<typeof resolveMemoryBackendConfig>;
+
+const resolveComparablePath = (value: string, workspaceDir = "/workspace/root"): string =>
+  path.isAbsolute(value) ? path.resolve(value) : path.resolve(workspaceDir, value);
+
+const memoryFileEntry = (name: string): Dirent =>
+  ({
+    name,
+    isFile: () => true,
+    isSymbolicLink: () => false,
+  }) as Dirent;
+
+const withMemoryRootEntries = <T>(entries: Dirent[], test: () => T): T => {
+  const readdirSpy = vi
+    .spyOn(syncFs, "readdirSync")
+    .mockReturnValue(entries as unknown as ReturnType<typeof syncFs.readdirSync>);
+  try {
+    return test();
+  } finally {
+    readdirSpy.mockRestore();
+  }
+};
+
+const rootMemoryConfig = (workspaceDir: string): OpenClawConfig =>
+  ({
+    agents: {
+      defaults: { workspace: workspaceDir },
+      list: [{ id: "main", default: true, workspace: workspaceDir }],
+    },
+    memory: {
+      backend: "qmd",
+      qmd: {},
+    },
+  }) as OpenClawConfig;
+
+const collectionNames = (resolved: ResolvedMemoryBackendConfig): string[] =>
+  (resolved.qmd?.collections ?? []).map((collection) => collection.name).toSorted();
+
+function requireQmdConfig(
+  resolved: ResolvedMemoryBackendConfig,
+): NonNullable<ResolvedMemoryBackendConfig["qmd"]> {
+  if (!resolved.qmd) {
+    throw new Error("expected qmd memory backend config");
+  }
+  return resolved.qmd;
+}
+
+function requireQmdCollection(
+  resolved: ResolvedMemoryBackendConfig,
+  name: string,
+): NonNullable<ResolvedMemoryBackendConfig["qmd"]>["collections"][number] {
+  const collection = requireQmdConfig(resolved).collections.find(
+    (candidate) => candidate.name === name,
+  );
+  if (!collection) {
+    throw new Error(`expected qmd collection ${name}`);
+  }
+  return collection;
+}
+
+const customQmdCollections = (
+  resolved: ResolvedMemoryBackendConfig,
+): NonNullable<ResolvedMemoryBackendConfig["qmd"]>["collections"] =>
+  (resolved.qmd?.collections ?? []).filter((collection) => collection.kind === "custom");
+
+const customCollectionPaths = (resolved: ResolvedMemoryBackendConfig): string[] =>
+  customQmdCollections(resolved)
+    .map((collection) => collection.path)
+    .toSorted();
+
+let fixtureRoot: string;
+let fixtureId = 0;
+
+beforeAll(async () => {
+  fixtureRoot = await fs.mkdtemp(path.join(os.tmpdir(), "qmd-backend-config-"));
+});
+
+afterAll(async () => {
+  await fs.rm(fixtureRoot, { recursive: true, force: true });
+});
+
+async function createFixtureDir(name: string): Promise<string> {
+  const dir = path.join(fixtureRoot, `${name}-${fixtureId++}`);
+  await fs.mkdir(dir, { recursive: true });
+  return dir;
+}
+
+describe("resolveMemoryBackendConfig", () => {
+  it("defaults to builtin backend when config missing", () => {
+    const cfg = { agents: { defaults: { workspace: "/tmp/memory-test" } } } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    expect(resolved.backend).toBe("builtin");
+    expect(resolved.citations).toBe("auto");
+    expect(resolved.qmd).toBeUndefined();
+  });
+
+  it("resolves qmd backend with default collections", () => {
+    const cfg = {
+      agents: { defaults: { workspace: "/tmp/memory-test" } },
+      memory: {
+        backend: "qmd",
+        qmd: {},
+      },
+    } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    expect(resolved.backend).toBe("qmd");
+    const qmd = requireQmdConfig(resolved);
+    expect(qmd.collections.length).toBe(2);
+    expect(qmd.command).toBe("qmd");
+    expect(qmd.searchMode).toBe("search");
+    expect(qmd.update.intervalMs).toBe(300_000);
+    expect(qmd.update.debounceMs).toBe(15_000);
+    expect(qmd.update.onBoot).toBe(true);
+    expect(qmd.update.startup).toBe("off");
+    expect(qmd.update.startupDelayMs).toBe(120_000);
+    expect(qmd.update.waitForBootSync).toBe(false);
+    expect(qmd.update.embedIntervalMs).toBe(3_600_000);
+    expect(qmd.update.commandTimeoutMs).toBe(30_000);
+    expect(qmd.update.updateTimeoutMs).toBe(120_000);
+    expect(qmd.update.embedTimeoutMs).toBe(120_000);
+    expect(collectionNames(resolved)).toStrictEqual(["memory-dir-main", "memory-root-main"]);
+    expect(requireQmdCollection(resolved, "memory-root-main").pattern).toBe("MEMORY.md");
+  });
+
+  it("keeps uppercase MEMORY.md as the root pattern when only lowercase memory.md exists", () => {
+    const workspaceDir = "/workspace/root";
+    withMemoryRootEntries([memoryFileEntry("memory.md")], () => {
+      const cfg = rootMemoryConfig(workspaceDir);
+      const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+      expect(requireQmdCollection(resolved, "memory-root-main").pattern).toBe("MEMORY.md");
+      expect(collectionNames(resolved)).toStrictEqual(["memory-dir-main", "memory-root-main"]);
+    });
+  });
+
+  it("prefers MEMORY.md over legacy memory.md when both root files exist", () => {
+    const workspaceDir = "/workspace/root";
+    withMemoryRootEntries([memoryFileEntry("MEMORY.md"), memoryFileEntry("memory.md")], () => {
+      const cfg = rootMemoryConfig(workspaceDir);
+      const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+      expect(requireQmdCollection(resolved, "memory-root-main").pattern).toBe("MEMORY.md");
+      expect(collectionNames(resolved)).toStrictEqual(["memory-dir-main", "memory-root-main"]);
+    });
+  });
+
+  it("parses quoted qmd command paths", () => {
+    const cfg = {
+      agents: { defaults: { workspace: "/tmp/memory-test" } },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          command: '"/Applications/QMD Tools/qmd" --flag',
+        },
+      },
+    } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    expect(requireQmdConfig(resolved).command).toBe("/Applications/QMD Tools/qmd");
+  });
+
+  it("preserves unquoted Windows absolute qmd command paths", () => {
+    const command = String.raw`C:\Users\penny\AppData\Roaming\npm\node_modules\@tobilu\qmd\dist\cli\qmd.js`;
+    const cfg = {
+      agents: { defaults: { workspace: String.raw`C:\Users\penny\.openclaw\workspace` } },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          command,
+        },
+      },
+    } as OpenClawConfig;
+
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+
+    expect(requireQmdConfig(resolved).command).toBe(command);
+  });
+
+  it("preserves Windows wrapper command paths before extra args", () => {
+    const cfg = {
+      agents: { defaults: { workspace: String.raw`C:\Users\penny\.openclaw\workspace` } },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          command: String.raw`C:\Program Files\qmd\qmd.cmd --json`,
+        },
+      },
+    } as OpenClawConfig;
+
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+
+    expect(requireQmdConfig(resolved).command).toBe(String.raw`C:\Program Files\qmd\qmd.cmd`);
+  });
+
+  it("preserves unquoted UNC qmd command paths", () => {
+    const command = String.raw`\\fileserver\tools\qmd\dist\cli\qmd.js`;
+    const cfg = {
+      agents: { defaults: { workspace: String.raw`C:\Users\penny\.openclaw\workspace` } },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          command,
+        },
+      },
+    } as OpenClawConfig;
+
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+
+    expect(requireQmdConfig(resolved).command).toBe(command);
+  });
+
+  it("resolves custom paths relative to workspace", () => {
+    const cfg = {
+      agents: {
+        defaults: { workspace: "/workspace/root" },
+        list: [{ id: "main", workspace: "/workspace/root" }],
+      },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          paths: [
+            {
+              path: "notes",
+              name: "custom-notes",
+              pattern: "**/*.md",
+            },
+          ],
+        },
+      },
+    } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const custom = requireQmdConfig(resolved).collections.find((c) =>
+      c.name.startsWith("custom-notes"),
+    );
+    if (!custom) {
+      throw new Error("expected custom-notes qmd collection");
+    }
+    expect(custom.path).toBe(path.resolve("/workspace/root", "notes"));
+  });
+
+  it("normalizes direct file qmd paths to escaped exact-file patterns", async () => {
+    const workspaceDir = await createFixtureDir("direct-file-path");
+    const notesPath = path.join(workspaceDir, "notes{a,b}[1].md");
+    await fs.writeFile(notesPath, "# Notes\n", "utf8");
+
+    const cfg = {
+      agents: {
+        defaults: { workspace: workspaceDir },
+        list: [{ id: "main", workspace: workspaceDir }],
+      },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          paths: [{ path: "notes{a,b}[1].md", name: "direct-note", pattern: "**/*.md" }],
+        },
+      },
+    } as OpenClawConfig;
+
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const custom = resolved.qmd?.collections.find((c) => c.name.startsWith("direct-note"));
+    expect(custom).toMatchObject({
+      path: workspaceDir,
+      pattern: String.raw`notes\{a,b\}\[1\].md`,
+    });
+  });
+
+  it("scopes qmd collection names per agent", () => {
+    const cfg = {
+      agents: {
+        defaults: { workspace: "/workspace/root" },
+        list: [
+          { id: "main", default: true, workspace: "/workspace/root" },
+          { id: "dev", workspace: "/workspace/dev" },
+        ],
+      },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          includeDefaultMemory: true,
+          paths: [{ path: "notes", name: "workspace", pattern: "**/*.md" }],
+        },
+      },
+    } as OpenClawConfig;
+    const mainResolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const devResolved = resolveMemoryBackendConfig({ cfg, agentId: "dev" });
+    const mainNames = collectionNames(mainResolved);
+    const devNames = collectionNames(devResolved);
+    expect(mainNames).toStrictEqual(["memory-dir-main", "memory-root-main", "workspace-main"]);
+    expect(devNames).toStrictEqual(["memory-dir-dev", "memory-root-dev", "workspace-dev"]);
+  });
+
+  it("merges default and per-agent qmd extra collections", () => {
+    const cfg = {
+      agents: {
+        defaults: {
+          workspace: "/workspace/root",
+          memorySearch: {
+            qmd: {
+              extraCollections: [
+                {
+                  path: "/shared/team-notes",
+                  name: "team-notes",
+                  pattern: "**/*.md",
+                },
+              ],
+            },
+          },
+        },
+        list: [
+          {
+            id: "main",
+            default: true,
+            workspace: "/workspace/root",
+            memorySearch: {
+              qmd: {
+                extraCollections: [
+                  {
+                    path: "notes",
+                    name: "notes",
+                    pattern: "**/*.md",
+                  },
+                ],
+              },
+            },
+          },
+        ],
+      },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          includeDefaultMemory: false,
+        },
+      },
+    } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const names = collectionNames(resolved);
+    expect(names).toStrictEqual(["notes-main", "team-notes"]);
+  });
+
+  it("preserves explicit custom collection names for paths outside the workspace", () => {
+    const cfg = {
+      agents: {
+        defaults: { workspace: "/workspace/root" },
+        list: [
+          { id: "main", default: true, workspace: "/workspace/root" },
+          { id: "dev", workspace: "/workspace/dev" },
+        ],
+      },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          includeDefaultMemory: true,
+          paths: [{ path: "/shared/notion-mirror", name: "notion-mirror", pattern: "**/*.md" }],
+        },
+      },
+    } as OpenClawConfig;
+    const mainResolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const devResolved = resolveMemoryBackendConfig({ cfg, agentId: "dev" });
+    const mainNames = collectionNames(mainResolved);
+    const devNames = collectionNames(devResolved);
+    expect(mainNames).toStrictEqual(["memory-dir-main", "memory-root-main", "notion-mirror"]);
+    expect(devNames).toStrictEqual(["memory-dir-dev", "memory-root-dev", "notion-mirror"]);
+  });
+
+  it("keeps symlinked workspace paths agent-scoped when deciding custom collection names", async () => {
+    const tmpRoot = await createFixtureDir("symlinked-workspace");
+    const workspaceDir = path.join(tmpRoot, "workspace");
+    const workspaceAliasDir = path.join(tmpRoot, "workspace-alias");
+    await fs.mkdir(workspaceDir, { recursive: true });
+    await fs.symlink(workspaceDir, workspaceAliasDir);
+    const cfg = {
+      agents: {
+        defaults: { workspace: workspaceDir },
+        list: [{ id: "main", default: true, workspace: workspaceDir }],
+      },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          includeDefaultMemory: false,
+          paths: [{ path: workspaceAliasDir, name: "workspace", pattern: "**/*.md" }],
+        },
+      },
+    } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const names = collectionNames(resolved);
+    expect(names).toStrictEqual(["workspace-main"]);
+  });
+
+  it("keeps unresolved child paths under a symlinked workspace agent-scoped", async () => {
+    const tmpRoot = await createFixtureDir("symlinked-child");
+    const realRootDir = path.join(tmpRoot, "real-root");
+    const aliasRootDir = path.join(tmpRoot, "alias-root");
+    const workspaceDir = path.join(realRootDir, "workspace");
+    const workspaceAliasDir = path.join(aliasRootDir, "workspace");
+    await fs.mkdir(workspaceDir, { recursive: true });
+    await fs.symlink(realRootDir, aliasRootDir);
+    const cfg = {
+      agents: {
+        defaults: { workspace: workspaceDir },
+        list: [{ id: "main", default: true, workspace: workspaceDir }],
+      },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          includeDefaultMemory: false,
+          paths: [
+            { path: path.join(workspaceAliasDir, "notes"), name: "notes", pattern: "**/*.md" },
+          ],
+        },
+      },
+    } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const names = collectionNames(resolved);
+    expect(names).toStrictEqual(["notes-main"]);
+  });
+
+  it("resolves qmd update timeout overrides", () => {
+    const cfg = {
+      agents: { defaults: { workspace: "/tmp/memory-test" } },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          update: {
+            waitForBootSync: true,
+            commandTimeoutMs: 12_000,
+            updateTimeoutMs: 480_000,
+            embedTimeoutMs: 360_000,
+          },
+        },
+      },
+    } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const update = requireQmdConfig(resolved).update;
+    expect(update.waitForBootSync).toBe(true);
+    expect(update.commandTimeoutMs).toBe(12_000);
+    expect(update.updateTimeoutMs).toBe(480_000);
+    expect(update.embedTimeoutMs).toBe(360_000);
+  });
+
+  it("keeps sub-unit positive qmd numeric overrides usable", () => {
+    const cfg = {
+      agents: { defaults: { workspace: "/tmp/memory-test" } },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          sessions: {
+            enabled: true,
+            retentionDays: 0.5,
+          },
+          update: {
+            commandTimeoutMs: 0.5,
+            updateTimeoutMs: 0.5,
+            embedTimeoutMs: 0.5,
+          },
+          limits: {
+            maxResults: 0.5,
+            maxSnippetChars: 0.5,
+            maxInjectedChars: 0.5,
+            timeoutMs: 0.5,
+          },
+        },
+      },
+    } as OpenClawConfig;
+
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const qmd = requireQmdConfig(resolved);
+
+    expect(qmd.sessions.retentionDays).toBe(1);
+    expect(qmd.update.commandTimeoutMs).toBe(1);
+    expect(qmd.update.updateTimeoutMs).toBe(1);
+    expect(qmd.update.embedTimeoutMs).toBe(1);
+    expect(qmd.limits).toMatchObject({
+      maxResults: 1,
+      maxSnippetChars: 1,
+      maxInjectedChars: 1,
+      timeoutMs: 1,
+    });
+  });
+
+  it("falls back for non-finite qmd numeric overrides", () => {
+    const cfg = {
+      agents: { defaults: { workspace: "/tmp/memory-test" } },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          sessions: {
+            enabled: true,
+            retentionDays: Number.NaN,
+          },
+          update: {
+            commandTimeoutMs: Number.POSITIVE_INFINITY,
+            updateTimeoutMs: Number.NaN,
+            embedTimeoutMs: Number.NEGATIVE_INFINITY,
+          },
+          limits: {
+            maxResults: Number.NaN,
+            maxSnippetChars: Number.POSITIVE_INFINITY,
+            maxInjectedChars: Number.NEGATIVE_INFINITY,
+            timeoutMs: Number.NaN,
+          },
+        },
+      },
+    } as OpenClawConfig;
+
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const qmd = requireQmdConfig(resolved);
+
+    expect(qmd.sessions.retentionDays).toBeUndefined();
+    expect(qmd.update.commandTimeoutMs).toBe(30_000);
+    expect(qmd.update.updateTimeoutMs).toBe(120_000);
+    expect(qmd.update.embedTimeoutMs).toBe(120_000);
+    expect(qmd.limits).toMatchObject({
+      maxResults: 4,
+      maxSnippetChars: 450,
+      maxInjectedChars: 2_200,
+      timeoutMs: 4_000,
+    });
+  });
+
+  it("resolves qmd startup refresh overrides", () => {
+    const cfg = {
+      agents: { defaults: { workspace: "/tmp/memory-test" } },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          update: {
+            startup: "idle",
+            startupDelayMs: 45_000,
+          },
+        },
+      },
+    } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const update = requireQmdConfig(resolved).update;
+    expect(update.startup).toBe("idle");
+    expect(update.startupDelayMs).toBe(45_000);
+    expect(update.onBoot).toBe(true);
+  });
+
+  it("resolves qmd search mode override", () => {
+    const cfg = {
+      agents: { defaults: { workspace: "/tmp/memory-test" } },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          searchMode: "vsearch",
+        },
+      },
+    } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    expect(requireQmdConfig(resolved).searchMode).toBe("vsearch");
+  });
+
+  it("resolves qmd rerank override", () => {
+    const cfg = {
+      agents: { defaults: { workspace: "/tmp/memory-test" } },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          searchMode: "query",
+          rerank: false,
+        },
+      },
+    } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const qmd = requireQmdConfig(resolved);
+    expect(qmd.searchMode).toBe("query");
+    expect(qmd.rerank).toBe(false);
+  });
+
+  it("resolves qmd mcporter search tool override", () => {
+    const cfg = {
+      agents: { defaults: { workspace: "/tmp/memory-test" } },
+      memory: {
+        backend: "qmd",
+        qmd: {
+          searchMode: "query",
+          searchTool: " hybrid_search ",
+        },
+      },
+    } as OpenClawConfig;
+    const resolved = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const qmd = requireQmdConfig(resolved);
+    expect(qmd.searchMode).toBe("query");
+    expect(qmd.searchTool).toBe("hybrid_search");
+  });
+});
+
+describe("memorySearch.extraPaths integration", () => {
+  it("maps agents.defaults.memorySearch.extraPaths to QMD collections", () => {
+    const cfg = {
+      memory: { backend: "qmd" },
+      agents: {
+        defaults: {
+          workspace: "/workspace/root",
+          memorySearch: {
+            extraPaths: ["/home/user/docs", "/home/user/vault"],
+          },
+        },
+      },
+    } as OpenClawConfig;
+    const result = resolveMemoryBackendConfig({ cfg, agentId: "test-agent" });
+    expect(result.backend).toBe("qmd");
+    const paths = customCollectionPaths(result);
+    expect(paths).toStrictEqual([
+      resolveComparablePath("/home/user/docs"),
+      resolveComparablePath("/home/user/vault"),
+    ]);
+  });
+
+  it("merges default and per-agent memorySearch.extraPaths for QMD collections", () => {
+    const cfg = {
+      memory: { backend: "qmd" },
+      agents: {
+        defaults: {
+          workspace: "/workspace/root",
+          memorySearch: {
+            extraPaths: ["/default/path"],
+          },
+        },
+        list: [
+          {
+            id: "my-agent",
+            memorySearch: {
+              extraPaths: ["/agent/specific/path"],
+            },
+          },
+        ],
+      },
+    } as OpenClawConfig;
+    const result = resolveMemoryBackendConfig({ cfg, agentId: "my-agent" });
+    expect(result.backend).toBe("qmd");
+    const paths = customCollectionPaths(result);
+    expect(paths).toStrictEqual([
+      resolveComparablePath("/agent/specific/path"),
+      resolveComparablePath("/default/path"),
+    ]);
+  });
+
+  it("falls back to defaults when agent has no overrides", () => {
+    const cfg = {
+      memory: { backend: "qmd" },
+      agents: {
+        defaults: {
+          workspace: "/workspace/root",
+          memorySearch: {
+            extraPaths: ["/default/path"],
+          },
+        },
+        list: [
+          {
+            id: "other-agent",
+            memorySearch: {
+              extraPaths: ["/other/path"],
+            },
+          },
+        ],
+      },
+    } as OpenClawConfig;
+    const result = resolveMemoryBackendConfig({ cfg, agentId: "my-agent" });
+    expect(result.backend).toBe("qmd");
+    const paths = customCollectionPaths(result);
+    expect(paths).toStrictEqual([resolveComparablePath("/default/path")]);
+  });
+
+  it("deduplicates merged memorySearch.extraPaths for QMD collections", () => {
+    const cfg = {
+      memory: { backend: "qmd" },
+      agents: {
+        defaults: {
+          workspace: "/workspace/root",
+          memorySearch: {
+            extraPaths: ["/shared/path", " /shared/path "],
+          },
+        },
+        list: [
+          {
+            id: "my-agent",
+            memorySearch: {
+              extraPaths: ["/shared/path", "/agent-only"],
+            },
+          },
+        ],
+      },
+    } as OpenClawConfig;
+
+    const result = resolveMemoryBackendConfig({ cfg, agentId: "my-agent" });
+    const paths = customCollectionPaths(result);
+
+    expect(paths).toStrictEqual([
+      resolveComparablePath("/agent-only"),
+      resolveComparablePath("/shared/path"),
+    ]);
+  });
+
+  it("keeps unnamed extra paths agent-scoped even when they resolve outside the workspace", () => {
+    const cfg = {
+      memory: { backend: "qmd" },
+      agents: {
+        defaults: {
+          workspace: "/workspace/root",
+          memorySearch: {
+            extraPaths: ["/shared/path"],
+          },
+        },
+      },
+    } as OpenClawConfig;
+    const result = resolveMemoryBackendConfig({ cfg, agentId: "my-agent" });
+    expect(customQmdCollections(result).map((collection) => collection.name)).toStrictEqual([
+      "custom-1-my-agent",
+    ]);
+  });
+
+  it("matches per-agent memorySearch.extraPaths using normalized agent ids", () => {
+    const cfg = {
+      memory: { backend: "qmd" },
+      agents: {
+        defaults: {
+          workspace: "/workspace/root",
+        },
+        list: [
+          {
+            id: "My-Agent",
+            memorySearch: {
+              extraPaths: ["/agent/mixed-case"],
+            },
+          },
+        ],
+      },
+    } as OpenClawConfig;
+
+    const result = resolveMemoryBackendConfig({ cfg, agentId: "my-agent" });
+
+    expect(customCollectionPaths(result)).toStrictEqual([
+      resolveComparablePath("/agent/mixed-case"),
+    ]);
+  });
+
+  it("deduplicates identical roots shared by memory.qmd.paths and memorySearch.extraPaths", () => {
+    const cfg = {
+      memory: {
+        backend: "qmd",
+        qmd: {
+          paths: [{ path: "docs", pattern: "**/*.md", name: "workspace-docs" }],
+        },
+      },
+      agents: {
+        defaults: {
+          workspace: "/workspace/root",
+          memorySearch: {
+            extraPaths: ["./docs"],
+          },
+        },
+      },
+    } as OpenClawConfig;
+
+    const result = resolveMemoryBackendConfig({ cfg, agentId: "main" });
+    const docsCollections = customQmdCollections(result).filter(
+      (collection) =>
+        collection.path === resolveComparablePath("./docs") && collection.pattern === "**/*.md",
+    );
+
+    expect(docsCollections).toHaveLength(1);
+  });
+});

packages/memory-host-sdk/src/host/backend-config.ts

@@ -0,0 +1,510 @@
+// Memory Host SDK module implements backend config behavior.
+import fs from "node:fs";
+import path from "node:path";
+import {
+  CANONICAL_ROOT_MEMORY_FILENAME,
+  type MemoryBackend,
+  type MemoryCitationsMode,
+  type MemoryQmdConfig,
+  type MemoryQmdIndexPath,
+  type MemoryQmdMcporterConfig,
+  type MemoryQmdSearchMode,
+  type MemoryQmdStartupMode,
+  type OpenClawConfig,
+  parseDurationMs,
+  resolveAgentWorkspaceDir,
+  normalizeAgentId,
+  resolveUserPath,
+  type SessionSendPolicyConfig,
+  splitShellArgs,
+} from "./config-utils.js";
+import { isPathInside } from "./fs-utils.js";
+import {
+  normalizeLowercaseStringOrEmpty,
+  normalizeStringEntries,
+  uniqueStrings,
+} from "./string-utils.js";
+
+function escapeQmdExactFilePattern(fileName: string): string {
+  return fileName.replace(/[\\*?[\]{}()!+@]/g, "\\$&");
+}
+
+const WINDOWS_COMMAND_EXTENSION_RE =
+  /^((?:[A-Za-z]:[\\/]|\\\\[^\\/]+[\\/][^\\/]+[\\/]).*?\.(?:bat|cmd|cjs|exe|js|mjs|ps1))(?:\s+|$)/i;
+
+function resolveQmdCommand(rawCommand: string): string {
+  const trimmedCommand = rawCommand.trim();
+  const windowsCommand = resolveWindowsAbsoluteCommand(trimmedCommand);
+  if (windowsCommand) {
+    return windowsCommand;
+  }
+
+  const parsedCommand = splitShellArgs(trimmedCommand);
+  return parsedCommand?.[0] || trimmedCommand.split(/\s+/)[0] || "qmd";
+}
+
+function resolveWindowsAbsoluteCommand(rawCommand: string): string | undefined {
+  if (!path.win32.isAbsolute(rawCommand)) {
+    return undefined;
+  }
+
+  const extensionMatch = WINDOWS_COMMAND_EXTENSION_RE.exec(rawCommand);
+  if (extensionMatch) {
+    return extensionMatch[1];
+  }
+
+  const firstWhitespace = rawCommand.search(/\s/);
+  return firstWhitespace === -1 ? rawCommand : rawCommand.slice(0, firstWhitespace);
+}
+
+export type ResolvedMemoryBackendConfig = {
+  backend: MemoryBackend;
+  citations: MemoryCitationsMode;
+  qmd?: ResolvedQmdConfig;
+};
+
+export type ResolvedQmdCollection = {
+  name: string;
+  path: string;
+  pattern: string;
+  kind: "memory" | "custom" | "sessions";
+};
+
+export type ResolvedQmdUpdateConfig = {
+  intervalMs: number;
+  debounceMs: number;
+  onBoot: boolean;
+  startup: MemoryQmdStartupMode;
+  startupDelayMs: number;
+  waitForBootSync: boolean;
+  embedIntervalMs: number;
+  commandTimeoutMs: number;
+  updateTimeoutMs: number;
+  embedTimeoutMs: number;
+};
+
+export type ResolvedQmdLimitsConfig = {
+  maxResults: number;
+  maxSnippetChars: number;
+  maxInjectedChars: number;
+  timeoutMs: number;
+};
+
+export type ResolvedQmdSessionConfig = {
+  enabled: boolean;
+  exportDir?: string;
+  retentionDays?: number;
+};
+
+export type ResolvedQmdMcporterConfig = {
+  enabled: boolean;
+  serverName: string;
+  startDaemon: boolean;
+};
+
+export type ResolvedQmdConfig = {
+  command: string;
+  mcporter: ResolvedQmdMcporterConfig;
+  searchMode: MemoryQmdSearchMode;
+  rerank?: boolean;
+  searchTool?: string;
+  collections: ResolvedQmdCollection[];
+  sessions: ResolvedQmdSessionConfig;
+  update: ResolvedQmdUpdateConfig;
+  limits: ResolvedQmdLimitsConfig;
+  includeDefaultMemory: boolean;
+  scope?: SessionSendPolicyConfig;
+};
+
+const DEFAULT_BACKEND: MemoryBackend = "builtin";
+const DEFAULT_CITATIONS: MemoryCitationsMode = "auto";
+const DEFAULT_QMD_INTERVAL = "5m";
+const DEFAULT_QMD_DEBOUNCE_MS = 15_000;
+const DEFAULT_QMD_TIMEOUT_MS = 4_000;
+// Defaulting to `query` can be extremely slow on CPU-only systems (query expansion + rerank).
+// Prefer a faster mode for interactive use; users can opt into `query` for best recall.
+const DEFAULT_QMD_SEARCH_MODE: MemoryQmdSearchMode = "search";
+const DEFAULT_QMD_STARTUP: MemoryQmdStartupMode = "off";
+const DEFAULT_QMD_STARTUP_DELAY_MS = 120_000;
+const DEFAULT_QMD_EMBED_INTERVAL = "60m";
+const DEFAULT_QMD_COMMAND_TIMEOUT_MS = 30_000;
+const DEFAULT_QMD_UPDATE_TIMEOUT_MS = 120_000;
+const DEFAULT_QMD_EMBED_TIMEOUT_MS = 120_000;
+const DEFAULT_QMD_LIMITS: ResolvedQmdLimitsConfig = {
+  maxResults: 4,
+  maxSnippetChars: 450,
+  maxInjectedChars: 2_200,
+  timeoutMs: DEFAULT_QMD_TIMEOUT_MS,
+};
+const DEFAULT_QMD_MCPORTER: ResolvedQmdMcporterConfig = {
+  enabled: false,
+  serverName: "qmd",
+  startDaemon: true,
+};
+
+const DEFAULT_QMD_SCOPE: SessionSendPolicyConfig = {
+  default: "deny",
+  rules: [
+    {
+      action: "allow",
+      match: { chatType: "direct" },
+    },
+  ],
+};
+
+function sanitizeName(input: string): string {
+  const lower = normalizeLowercaseStringOrEmpty(input).replace(/[^a-z0-9-]+/g, "-");
+  const trimmed = lower.replace(/^-+|-+$/g, "");
+  return trimmed || "collection";
+}
+
+function scopeCollectionBase(base: string, agentId: string): string {
+  return `${base}-${sanitizeName(agentId)}`;
+}
+
+function canonicalizePathForContainment(rawPath: string): string {
+  const resolved = path.resolve(rawPath);
+  let current = resolved;
+  const suffix: string[] = [];
+  while (true) {
+    try {
+      const canonical = path.normalize(fs.realpathSync.native(current));
+      return path.normalize(path.join(canonical, ...suffix));
+    } catch {
+      const parent = path.dirname(current);
+      if (parent === current) {
+        return path.normalize(resolved);
+      }
+      suffix.unshift(path.basename(current));
+      current = parent;
+    }
+  }
+}
+
+function isPathInsideRoot(candidatePath: string, rootPath: string): boolean {
+  return isPathInside(
+    canonicalizePathForContainment(rootPath),
+    canonicalizePathForContainment(candidatePath),
+  );
+}
+
+function ensureUniqueName(base: string, existing: Set<string>): string {
+  const name = sanitizeName(base);
+  if (!existing.has(name)) {
+    existing.add(name);
+    return name;
+  }
+  let suffix = 2;
+  while (existing.has(`${name}-${suffix}`)) {
+    suffix += 1;
+  }
+  const unique = `${name}-${suffix}`;
+  existing.add(unique);
+  return unique;
+}
+
+function resolvePath(raw: string, workspaceDir: string): string {
+  const trimmed = raw.trim();
+  if (!trimmed) {
+    throw new Error("path required");
+  }
+  if (trimmed.startsWith("~") || path.isAbsolute(trimmed)) {
+    return path.normalize(resolveUserPath(trimmed));
+  }
+  return path.normalize(path.resolve(workspaceDir, trimmed));
+}
+
+function resolveIntervalMs(raw: string | undefined): number {
+  const value = raw?.trim();
+  if (!value) {
+    return parseDurationMs(DEFAULT_QMD_INTERVAL, { defaultUnit: "m" });
+  }
+  try {
+    return parseDurationMs(value, { defaultUnit: "m" });
+  } catch {
+    return parseDurationMs(DEFAULT_QMD_INTERVAL, { defaultUnit: "m" });
+  }
+}
+
+function resolveEmbedIntervalMs(raw: string | undefined): number {
+  const value = raw?.trim();
+  if (!value) {
+    return parseDurationMs(DEFAULT_QMD_EMBED_INTERVAL, { defaultUnit: "m" });
+  }
+  try {
+    return parseDurationMs(value, { defaultUnit: "m" });
+  } catch {
+    return parseDurationMs(DEFAULT_QMD_EMBED_INTERVAL, { defaultUnit: "m" });
+  }
+}
+
+function resolveDebounceMs(raw: number | undefined): number {
+  if (typeof raw === "number" && Number.isFinite(raw) && raw >= 0) {
+    return Math.floor(raw);
+  }
+  return DEFAULT_QMD_DEBOUNCE_MS;
+}
+
+function resolveTimeoutMs(raw: number | undefined, fallback: number): number {
+  return resolvePositiveIntegerConfig(raw, fallback);
+}
+
+function resolvePositiveIntegerConfig(raw: number | undefined, fallback: number): number;
+function resolvePositiveIntegerConfig(raw: number | undefined): number | undefined;
+function resolvePositiveIntegerConfig(
+  raw: number | undefined,
+  fallback?: number,
+): number | undefined {
+  if (typeof raw !== "number" || !Number.isFinite(raw) || raw <= 0) {
+    return fallback;
+  }
+  return Math.max(1, Math.floor(raw));
+}
+
+function resolveStartupMode(raw: MemoryQmdConfig["update"]): MemoryQmdStartupMode {
+  const value = raw?.startup;
+  if (value === "idle" || value === "immediate" || value === "off") {
+    return value;
+  }
+  return DEFAULT_QMD_STARTUP;
+}
+
+function resolveStartupDelayMs(raw: number | undefined): number {
+  if (typeof raw === "number" && Number.isFinite(raw) && raw >= 0) {
+    return Math.floor(raw);
+  }
+  return DEFAULT_QMD_STARTUP_DELAY_MS;
+}
+
+function resolveLimits(raw?: MemoryQmdConfig["limits"]): ResolvedQmdLimitsConfig {
+  return {
+    maxResults: resolvePositiveIntegerConfig(raw?.maxResults, DEFAULT_QMD_LIMITS.maxResults),
+    maxSnippetChars: resolvePositiveIntegerConfig(
+      raw?.maxSnippetChars,
+      DEFAULT_QMD_LIMITS.maxSnippetChars,
+    ),
+    maxInjectedChars: resolvePositiveIntegerConfig(
+      raw?.maxInjectedChars,
+      DEFAULT_QMD_LIMITS.maxInjectedChars,
+    ),
+    timeoutMs: resolvePositiveIntegerConfig(raw?.timeoutMs, DEFAULT_QMD_LIMITS.timeoutMs),
+  };
+}
+
+function resolveSearchMode(raw?: MemoryQmdConfig["searchMode"]): MemoryQmdSearchMode {
+  if (raw === "search" || raw === "vsearch" || raw === "query") {
+    return raw;
+  }
+  return DEFAULT_QMD_SEARCH_MODE;
+}
+
+function resolveSearchTool(raw?: MemoryQmdConfig["searchTool"]): string | undefined {
+  const value = raw?.trim();
+  return value ? value : undefined;
+}
+
+function resolveSessionConfig(
+  cfg: MemoryQmdConfig["sessions"],
+  workspaceDir: string,
+): ResolvedQmdSessionConfig {
+  const enabled = Boolean(cfg?.enabled);
+  const exportDirRaw = cfg?.exportDir?.trim();
+  const exportDir = exportDirRaw ? resolvePath(exportDirRaw, workspaceDir) : undefined;
+  const retentionDays = resolvePositiveIntegerConfig(cfg?.retentionDays);
+  return {
+    enabled,
+    exportDir,
+    retentionDays,
+  };
+}
+
+function resolveCustomPaths(
+  rawPaths: MemoryQmdIndexPath[] | undefined,
+  workspaceDir: string,
+  existing: Set<string>,
+  agentId: string,
+): ResolvedQmdCollection[] {
+  if (!rawPaths?.length) {
+    return [];
+  }
+  const collections: ResolvedQmdCollection[] = [];
+  const seenRoots = new Set<string>();
+  rawPaths.forEach((entry, index) => {
+    const trimmedPath = entry?.path?.trim();
+    if (!trimmedPath) {
+      return;
+    }
+    let resolved: string;
+    let collectionPath: string;
+    try {
+      resolved = resolvePath(trimmedPath, workspaceDir);
+    } catch {
+      return;
+    }
+    collectionPath = resolved;
+    let pattern = entry.pattern?.trim() || "**/*.md";
+    try {
+      const stat = fs.statSync(resolved);
+      if (stat.isFile()) {
+        // When the configured path points directly to a file, normalize into a
+        // parent-directory collection with an exact-filename pattern, regardless
+        // of any user-supplied glob (a glob does not apply to a single file).
+        collectionPath = path.dirname(resolved);
+        pattern = escapeQmdExactFilePattern(path.basename(resolved));
+      }
+    } catch {
+      // not a file or can't stat, use as-is
+    }
+    const dedupeKey = `${collectionPath}\u0000${pattern}`;
+    if (seenRoots.has(dedupeKey)) {
+      return;
+    }
+    seenRoots.add(dedupeKey);
+    const explicitName = entry.name?.trim();
+    const baseName =
+      explicitName && !isPathInsideRoot(collectionPath, workspaceDir)
+        ? explicitName
+        : scopeCollectionBase(explicitName || `custom-${index + 1}`, agentId);
+    const name = ensureUniqueName(baseName, existing);
+    collections.push({
+      name,
+      path: collectionPath,
+      pattern,
+      kind: "custom",
+    });
+  });
+  return collections;
+}
+
+function resolveMcporterConfig(raw?: MemoryQmdMcporterConfig): ResolvedQmdMcporterConfig {
+  const parsed: ResolvedQmdMcporterConfig = { ...DEFAULT_QMD_MCPORTER };
+  if (!raw) {
+    return parsed;
+  }
+  if (raw.enabled !== undefined) {
+    parsed.enabled = raw.enabled;
+  }
+  if (typeof raw.serverName === "string" && raw.serverName.trim()) {
+    parsed.serverName = raw.serverName.trim();
+  }
+  if (raw.startDaemon !== undefined) {
+    parsed.startDaemon = raw.startDaemon;
+  }
+  // When enabled, default startDaemon to true.
+  if (parsed.enabled && raw.startDaemon === undefined) {
+    parsed.startDaemon = true;
+  }
+  return parsed;
+}
+
+function resolveDefaultCollections(
+  include: boolean,
+  workspaceDir: string,
+  existing: Set<string>,
+  agentId: string,
+): ResolvedQmdCollection[] {
+  if (!include) {
+    return [];
+  }
+  const entries: Array<{ path: string; pattern: string; base: string }> = [
+    { path: workspaceDir, pattern: CANONICAL_ROOT_MEMORY_FILENAME, base: "memory-root" },
+    { path: path.join(workspaceDir, "memory"), pattern: "**/*.md", base: "memory-dir" },
+  ];
+  return entries.map((entry) => ({
+    name: ensureUniqueName(scopeCollectionBase(entry.base, agentId), existing),
+    path: entry.path,
+    pattern: entry.pattern,
+    kind: "memory",
+  }));
+}
+
+export function resolveMemoryBackendConfig(params: {
+  cfg: OpenClawConfig;
+  agentId: string;
+}): ResolvedMemoryBackendConfig {
+  const normalizedAgentId = normalizeAgentId(params.agentId);
+  const backend = params.cfg.memory?.backend ?? DEFAULT_BACKEND;
+  const citations = params.cfg.memory?.citations ?? DEFAULT_CITATIONS;
+  if (backend !== "qmd") {
+    return { backend: "builtin", citations };
+  }
+
+  const workspaceDir = resolveAgentWorkspaceDir(params.cfg, normalizedAgentId);
+  const qmdCfg = params.cfg.memory?.qmd;
+  const includeDefaultMemory = qmdCfg?.includeDefaultMemory !== false;
+  const nameSet = new Set<string>();
+  const agentEntry = params.cfg.agents?.list?.find(
+    (entry) => normalizeAgentId(entry?.id) === normalizedAgentId,
+  );
+  const mergedExtraPaths = normalizeStringEntries(
+    [
+      ...(params.cfg.agents?.defaults?.memorySearch?.extraPaths ?? []),
+      ...(agentEntry?.memorySearch?.extraPaths ?? []),
+    ].filter((value): value is string => typeof value === "string"),
+  );
+  const dedupedExtraPaths = uniqueStrings(mergedExtraPaths);
+  const searchExtraPaths = dedupedExtraPaths.map(
+    (pathValue): { path: string; pattern?: string; name?: string } => ({ path: pathValue }),
+  );
+  const mergedExtraCollections = [
+    ...(params.cfg.agents?.defaults?.memorySearch?.qmd?.extraCollections ?? []),
+    ...(agentEntry?.memorySearch?.qmd?.extraCollections ?? []),
+  ].filter(
+    (value): value is MemoryQmdIndexPath =>
+      value !== null && typeof value === "object" && typeof value.path === "string",
+  );
+
+  // Combine QMD-specific paths with extraPaths and per-agent cross-agent collections.
+  const allQmdPaths: MemoryQmdIndexPath[] = [
+    ...(qmdCfg?.paths ?? []),
+    ...searchExtraPaths,
+    ...mergedExtraCollections,
+  ];
+
+  const collections = [
+    ...resolveDefaultCollections(includeDefaultMemory, workspaceDir, nameSet, normalizedAgentId),
+    ...resolveCustomPaths(allQmdPaths, workspaceDir, nameSet, normalizedAgentId),
+  ];
+
+  const rawCommand = qmdCfg?.command?.trim() || "qmd";
+  const command = resolveQmdCommand(rawCommand);
+  const resolved: ResolvedQmdConfig = {
+    command,
+    mcporter: resolveMcporterConfig(qmdCfg?.mcporter),
+    searchMode: resolveSearchMode(qmdCfg?.searchMode),
+    rerank: qmdCfg?.rerank,
+    searchTool: resolveSearchTool(qmdCfg?.searchTool),
+    collections,
+    includeDefaultMemory,
+    sessions: resolveSessionConfig(qmdCfg?.sessions, workspaceDir),
+    update: {
+      intervalMs: resolveIntervalMs(qmdCfg?.update?.interval),
+      debounceMs: resolveDebounceMs(qmdCfg?.update?.debounceMs),
+      onBoot: qmdCfg?.update?.onBoot !== false,
+      startup: resolveStartupMode(qmdCfg?.update),
+      startupDelayMs: resolveStartupDelayMs(qmdCfg?.update?.startupDelayMs),
+      waitForBootSync: qmdCfg?.update?.waitForBootSync === true,
+      embedIntervalMs: resolveEmbedIntervalMs(qmdCfg?.update?.embedInterval),
+      commandTimeoutMs: resolveTimeoutMs(
+        qmdCfg?.update?.commandTimeoutMs,
+        DEFAULT_QMD_COMMAND_TIMEOUT_MS,
+      ),
+      updateTimeoutMs: resolveTimeoutMs(
+        qmdCfg?.update?.updateTimeoutMs,
+        DEFAULT_QMD_UPDATE_TIMEOUT_MS,
+      ),
+      embedTimeoutMs: resolveTimeoutMs(
+        qmdCfg?.update?.embedTimeoutMs,
+        DEFAULT_QMD_EMBED_TIMEOUT_MS,
+      ),
+    },
+    limits: resolveLimits(qmdCfg?.limits),
+    scope: qmdCfg?.scope ?? DEFAULT_QMD_SCOPE,
+  };
+
+  return {
+    backend: "qmd",
+    citations,
+    qmd: resolved,
+  };
+}

packages/memory-host-sdk/src/host/batch-error-utils.test.ts

@@ -0,0 +1,33 @@
+// Memory Host SDK tests cover batch error utils behavior.
+import { describe, expect, it } from "vitest";
+import { extractBatchErrorMessage, formatUnavailableBatchError } from "./batch-error-utils.js";
+
+describe("extractBatchErrorMessage", () => {
+  it("returns the first top-level error message", () => {
+    expect(
+      extractBatchErrorMessage([
+        { response: { body: { error: { message: "nested" } } } },
+        { error: { message: "top-level" } },
+      ]),
+    ).toBe("nested");
+  });
+
+  it("falls back to nested response error message", () => {
+    expect(
+      extractBatchErrorMessage([{ response: { body: { error: { message: "nested-only" } } } }, {}]),
+    ).toBe("nested-only");
+  });
+
+  it("accepts plain string response bodies", () => {
+    expect(extractBatchErrorMessage([{ response: { body: "provider plain-text error" } }])).toBe(
+      "provider plain-text error",
+    );
+  });
+});
+
+describe("formatUnavailableBatchError", () => {
+  it("formats errors and non-error values", () => {
+    expect(formatUnavailableBatchError(new Error("boom"))).toBe("error file unavailable: boom");
+    expect(formatUnavailableBatchError("unreachable")).toBe("error file unavailable: unreachable");
+  });
+});

packages/memory-host-sdk/src/host/batch-error-utils.ts

@@ -0,0 +1,40 @@
+// Memory Host SDK helper module supports batch error utils behavior.
+import { formatErrorMessage } from "./error-utils.js";
+
+// Extracts provider batch error text from output and unavailable error files.
+
+/** Minimal batch output line shape that can carry provider error messages. */
+type BatchOutputErrorLike = {
+  error?: { message?: string };
+  response?: {
+    body?:
+      | string
+      | {
+          error?: { message?: string };
+        };
+  };
+};
+
+/** Pull a nested response error message without assuming a fixed provider body shape. */
+function getResponseErrorMessage(line: BatchOutputErrorLike | undefined): string | undefined {
+  const body = line?.response?.body;
+  if (typeof body === "string") {
+    return body || undefined;
+  }
+  if (!body || typeof body !== "object") {
+    return undefined;
+  }
+  return typeof body.error?.message === "string" ? body.error.message : undefined;
+}
+
+/** Return the first useful error message from batch output lines. */
+export function extractBatchErrorMessage(lines: BatchOutputErrorLike[]): string | undefined {
+  const first = lines.find((line) => line.error?.message || getResponseErrorMessage(line));
+  return first?.error?.message ?? getResponseErrorMessage(first);
+}
+
+/** Format a failed error-file read without hiding the underlying read problem. */
+export function formatUnavailableBatchError(err: unknown): string | undefined {
+  const message = formatErrorMessage(err);
+  return message ? `error file unavailable: ${message}` : undefined;
+}

packages/memory-host-sdk/src/host/batch-http.test.ts

@@ -0,0 +1,114 @@
+// Memory Host SDK tests cover batch http behavior.
+import { beforeAll, beforeEach, describe, expect, it, vi } from "vitest";
+
+vi.mock("./post-json.js", () => ({
+  postJson: vi.fn(),
+}));
+
+type RetryOptions = {
+  attempts: number;
+  minDelayMs: number;
+  maxDelayMs: number;
+  shouldRetry: (err: unknown) => boolean;
+};
+
+type PostJsonParams = {
+  url?: unknown;
+  headers?: unknown;
+  body?: unknown;
+  errorPrefix?: unknown;
+  attachStatus?: unknown;
+};
+
+function requirePostJsonParams(
+  postJsonMock: ReturnType<typeof vi.mocked<typeof import("./post-json.js").postJson>>,
+): PostJsonParams {
+  const [call] = postJsonMock.mock.calls;
+  if (!call) {
+    throw new Error("expected postJson call");
+  }
+  const [params] = call;
+  if (typeof params !== "object" || params === null || Array.isArray(params)) {
+    throw new Error("expected postJson params to be an object");
+  }
+  return params;
+}
+
+function requireFirstRetryOptions(retryAsyncMock: ReturnType<typeof vi.fn>): RetryOptions {
+  const call = retryAsyncMock.mock.calls[0];
+  const options = call?.[1] as RetryOptions | undefined;
+  if (!options) {
+    throw new Error("expected retry options");
+  }
+  return options;
+}
+
+describe("postJsonWithRetry", () => {
+  let postJsonMock: ReturnType<typeof vi.mocked<typeof import("./post-json.js").postJson>>;
+  let postJsonWithRetry: typeof import("./batch-http.js").postJsonWithRetry;
+  let retryAsyncMock: ReturnType<typeof vi.fn>;
+
+  beforeAll(async () => {
+    ({ postJsonWithRetry } = await import("./batch-http.js"));
+    const postJsonModule = await import("./post-json.js");
+    postJsonMock = vi.mocked(postJsonModule.postJson);
+  });
+
+  beforeEach(() => {
+    vi.clearAllMocks();
+    retryAsyncMock = vi.fn(async (run: () => Promise<unknown>) => await run());
+  });
+
+  it("posts JSON and returns parsed response payload", async () => {
+    postJsonMock.mockImplementationOnce(async (params) => {
+      return await params.parse({ ok: true, ids: [1, 2] });
+    });
+
+    const result = await postJsonWithRetry<{ ok: boolean; ids: number[] }>({
+      url: "https://memory.example/v1/batch",
+      headers: { Authorization: "Bearer test" },
+      body: { chunks: ["a", "b"] },
+      errorPrefix: "memory batch failed",
+      retryImpl: retryAsyncMock as typeof import("./retry-utils.js").retryAsync,
+    });
+
+    expect(result).toEqual({ ok: true, ids: [1, 2] });
+    const postJsonParams = requirePostJsonParams(postJsonMock);
+    expect(postJsonParams.url).toBe("https://memory.example/v1/batch");
+    expect(postJsonParams.headers).toEqual({ Authorization: "Bearer test" });
+    expect(postJsonParams.body).toEqual({ chunks: ["a", "b"] });
+    expect(postJsonParams.errorPrefix).toBe("memory batch failed");
+    expect(postJsonParams.attachStatus).toBe(true);
+
+    const retryOptions = requireFirstRetryOptions(retryAsyncMock);
+    expect(retryOptions.attempts).toBe(3);
+    expect(retryOptions.minDelayMs).toBe(300);
+    expect(retryOptions.maxDelayMs).toBe(2000);
+    expect(retryOptions.shouldRetry({ status: 429 })).toBe(true);
+    expect(retryOptions.shouldRetry({ status: 503 })).toBe(true);
+    expect(retryOptions.shouldRetry({ status: 400 })).toBe(false);
+  });
+
+  it("attaches status to non-ok errors", async () => {
+    postJsonMock.mockRejectedValueOnce(
+      Object.assign(new Error("memory batch failed: 503 backend down"), { status: 503 }),
+    );
+
+    let error: unknown;
+    try {
+      await postJsonWithRetry({
+        url: "https://memory.example/v1/batch",
+        headers: {},
+        body: { chunks: [] },
+        errorPrefix: "memory batch failed",
+        retryImpl: retryAsyncMock as typeof import("./retry-utils.js").retryAsync,
+      });
+    } catch (caught) {
+      error = caught;
+    }
+
+    expect(error).toBeInstanceOf(Error);
+    expect((error as Error).message).toBe("memory batch failed: 503 backend down");
+    expect((error as { status?: unknown }).status).toBe(503);
+  });
+});

packages/memory-host-sdk/src/host/batch-http.ts

@@ -0,0 +1,43 @@
+// Memory Host SDK module implements batch http behavior.
+import { postJson } from "./post-json.js";
+import { retryAsync } from "./retry-utils.js";
+import type { SsrFPolicy } from "./ssrf-policy.js";
+
+// JSON POST helper for batch APIs with provider-style transient retry.
+
+/** POST JSON and retry provider 429/5xx failures with bounded backoff. */
+export async function postJsonWithRetry<T>(params: {
+  url: string;
+  headers: Record<string, string>;
+  ssrfPolicy?: SsrFPolicy;
+  fetchImpl?: typeof fetch;
+  retryImpl?: typeof retryAsync;
+  body: unknown;
+  errorPrefix: string;
+}): Promise<T> {
+  const retry = params.retryImpl ?? retryAsync;
+  return await retry(
+    async () => {
+      return await postJson<T>({
+        url: params.url,
+        headers: params.headers,
+        ssrfPolicy: params.ssrfPolicy,
+        fetchImpl: params.fetchImpl,
+        body: params.body,
+        errorPrefix: params.errorPrefix,
+        attachStatus: true,
+        parse: async (payload) => payload as T,
+      });
+    },
+    {
+      attempts: 3,
+      minDelayMs: 300,
+      maxDelayMs: 2000,
+      jitter: 0.2,
+      shouldRetry: (err) => {
+        const status = (err as { status?: number }).status;
+        return status === 429 || (typeof status === "number" && status >= 500);
+      },
+    },
+  );
+}

packages/memory-host-sdk/src/host/batch-output.test.ts

@@ -0,0 +1,83 @@
+// Memory Host SDK tests cover batch output behavior.
+import { describe, expect, it } from "vitest";
+import { applyEmbeddingBatchOutputLine } from "./batch-output.js";
+
+describe("applyEmbeddingBatchOutputLine", () => {
+  it("stores embedding for successful response", () => {
+    const remaining = new Set(["req-1"]);
+    const errors: string[] = [];
+    const byCustomId = new Map<string, number[]>();
+
+    applyEmbeddingBatchOutputLine({
+      line: {
+        custom_id: "req-1",
+        response: {
+          status_code: 200,
+          body: { data: [{ embedding: [0.1, 0.2] }] },
+        },
+      },
+      remaining,
+      errors,
+      byCustomId,
+    });
+
+    expect(remaining.has("req-1")).toBe(false);
+    expect(errors).toStrictEqual([]);
+    expect(byCustomId.get("req-1")).toEqual([0.1, 0.2]);
+  });
+
+  it("records provider error from line.error", () => {
+    const remaining = new Set(["req-2"]);
+    const errors: string[] = [];
+    const byCustomId = new Map<string, number[]>();
+
+    applyEmbeddingBatchOutputLine({
+      line: {
+        custom_id: "req-2",
+        error: { message: "provider failed" },
+      },
+      remaining,
+      errors,
+      byCustomId,
+    });
+
+    expect(remaining.has("req-2")).toBe(false);
+    expect(errors).toEqual(["req-2: provider failed"]);
+    expect(byCustomId.size).toBe(0);
+  });
+
+  it("records non-2xx response errors and empty embedding errors", () => {
+    const remaining = new Set(["req-3", "req-4"]);
+    const errors: string[] = [];
+    const byCustomId = new Map<string, number[]>();
+
+    applyEmbeddingBatchOutputLine({
+      line: {
+        custom_id: "req-3",
+        response: {
+          status_code: 500,
+          body: { error: { message: "internal" } },
+        },
+      },
+      remaining,
+      errors,
+      byCustomId,
+    });
+
+    applyEmbeddingBatchOutputLine({
+      line: {
+        custom_id: "req-4",
+        response: {
+          status_code: 200,
+          body: { data: [] },
+        },
+      },
+      remaining,
+      errors,
+      byCustomId,
+    });
+
+    expect(errors).toEqual(["req-3: internal", "req-4: empty embedding"]);
+    expect(byCustomId.size).toBe(0);
+  });
+});

packages/memory-host-sdk/src/host/batch-output.ts

@@ -0,0 +1,59 @@
+// Parses provider batch output lines into the custom-id embedding map.
+
+/** Minimal OpenAI-compatible embedding batch output line. */
+export type EmbeddingBatchOutputLine = {
+  custom_id?: string;
+  error?: { message?: string };
+  response?: {
+    status_code?: number;
+    body?:
+      | {
+          data?: Array<{
+            embedding?: number[];
+          }>;
+          error?: { message?: string };
+        }
+      | string;
+  };
+};
+
+/** Apply one output line, collecting errors and successful embeddings by custom id. */
+export function applyEmbeddingBatchOutputLine(params: {
+  line: EmbeddingBatchOutputLine;
+  remaining: Set<string>;
+  errors: string[];
+  byCustomId: Map<string, number[]>;
+}) {
+  const customId = params.line.custom_id;
+  if (!customId) {
+    return;
+  }
+  params.remaining.delete(customId);
+
+  const errorMessage = params.line.error?.message;
+  if (errorMessage) {
+    params.errors.push(`${customId}: ${errorMessage}`);
+    return;
+  }
+
+  const response = params.line.response;
+  const statusCode = response?.status_code ?? 0;
+  if (statusCode >= 400) {
+    const messageFromObject =
+      response?.body && typeof response.body === "object"
+        ? (response.body as { error?: { message?: string } }).error?.message
+        : undefined;
+    const messageFromString = typeof response?.body === "string" ? response.body : undefined;
+    params.errors.push(`${customId}: ${messageFromObject ?? messageFromString ?? "unknown error"}`);
+    return;
+  }
+
+  const data =
+    response?.body && typeof response.body === "object" ? (response.body.data ?? []) : [];
+  const embedding = data[0]?.embedding ?? [];
+  if (embedding.length === 0) {
+    params.errors.push(`${customId}: empty embedding`);
+    return;
+  }
+  params.byCustomId.set(customId, embedding);
+}

packages/memory-host-sdk/src/host/batch-provider-common.ts

@@ -0,0 +1,18 @@
+// Memory Host SDK helper module supports batch provider common behavior.
+import type { EmbeddingBatchOutputLine } from "./batch-output.js";
+
+// Common OpenAI-compatible batch shapes shared by remote embedding providers.
+
+/** Minimal provider batch status payload used by polling code. */
+export type EmbeddingBatchStatus = {
+  id?: string;
+  status?: string;
+  output_file_id?: string | null;
+  error_file_id?: string | null;
+};
+
+/** Provider output line after an embedding batch file is read. */
+export type ProviderBatchOutputLine = EmbeddingBatchOutputLine;
+
+/** OpenAI-compatible endpoint used inside embedding batch request lines. */
+export const EMBEDDING_BATCH_ENDPOINT = "/v1/embeddings";

packages/memory-host-sdk/src/host/batch-runner.test.ts

@@ -0,0 +1,157 @@
+// Memory Host SDK tests cover batch runner behavior.
+import { describe, expect, it, vi } from "vitest";
+import { MAX_SAFE_TIMEOUT_DELAY_MS } from "../../../gateway-client/src/timeouts.js";
+import { buildEmbeddingBatchGroupOptions, runEmbeddingBatchGroups } from "./batch-runner.js";
+
+const jsonlEncoder = new TextEncoder();
+
+function jsonlLineBytes(value: unknown): number {
+  return jsonlEncoder.encode(JSON.stringify(value)).byteLength;
+}
+
+describe("buildEmbeddingBatchGroupOptions", () => {
+  it("clamps oversized embedding batch poll intervals to the timeout budget", () => {
+    const options = buildEmbeddingBatchGroupOptions(
+      {
+        requests: ["request-1"],
+        wait: true,
+        pollIntervalMs: Number.MAX_SAFE_INTEGER,
+        timeoutMs: 60_000,
+        concurrency: 1,
+      },
+      {
+        maxRequests: 100,
+        debugLabel: "embedding batch submit",
+      },
+    );
+
+    expect(options.pollIntervalMs).toBe(60_000);
+  });
+
+  it("passes clamped poll intervals into batch group runners", async () => {
+    const runGroup = vi.fn(async () => {});
+
+    await runEmbeddingBatchGroups({
+      requests: ["request-1"],
+      maxRequests: 100,
+      wait: true,
+      pollIntervalMs: Number.MAX_SAFE_INTEGER,
+      timeoutMs: 60_000,
+      concurrency: 1,
+      debugLabel: "embedding batch submit",
+      runGroup,
+    });
+
+    expect(runGroup).toHaveBeenCalledWith(
+      expect.objectContaining({
+        pollIntervalMs: 60_000,
+        timeoutMs: 60_000,
+      }),
+    );
+  });
+
+  it("keeps timeout-safe oversized embedding batch poll intervals bounded", () => {
+    const options = buildEmbeddingBatchGroupOptions(
+      {
+        requests: ["request-1"],
+        wait: true,
+        pollIntervalMs: Number.MAX_SAFE_INTEGER,
+        timeoutMs: Number.MAX_SAFE_INTEGER,
+        concurrency: 1,
+      },
+      {
+        maxRequests: 100,
+        debugLabel: "embedding batch submit",
+      },
+    );
+
+    expect(options.pollIntervalMs).toBe(MAX_SAFE_TIMEOUT_DELAY_MS);
+  });
+
+  it("splits embedding batch groups by serialized JSONL bytes", async () => {
+    const requests = [
+      { id: "one", body: { input: "alpha" } },
+      { id: "two", body: { input: "βeta" } },
+      { id: "three", body: { input: "gamma" } },
+    ];
+    const maxJsonlBytes = jsonlLineBytes(requests[0]) + 1 + jsonlLineBytes(requests[1]);
+    const groups: string[][] = [];
+
+    await runEmbeddingBatchGroups({
+      requests,
+      maxRequests: 100,
+      maxJsonlBytes,
+      wait: true,
+      pollIntervalMs: 1000,
+      timeoutMs: 60_000,
+      concurrency: 1,
+      debugLabel: "embedding batch submit",
+      runGroup: async ({ group }) => {
+        groups.push(group.map((request) => request.id));
+      },
+    });
+
+    expect(groups).toEqual([["one", "two"], ["three"]]);
+  });
+
+  it("splits provider-rejected batch groups when the error is splittable", async () => {
+    const uploadTooLarge = new Error("batch upload failed: 413 payload too large");
+    const calls: string[][] = [];
+    const onSplitGroup = vi.fn();
+
+    await runEmbeddingBatchGroups({
+      requests: ["one", "two", "three", "four"],
+      maxRequests: 100,
+      wait: true,
+      pollIntervalMs: 1000,
+      timeoutMs: 60_000,
+      concurrency: 1,
+      debugLabel: "embedding batch submit",
+      shouldSplitGroupOnError: (error) => error === uploadTooLarge,
+      onSplitGroup,
+      runGroup: async ({ group }) => {
+        calls.push([...group]);
+        if (group.length === 4) {
+          throw uploadTooLarge;
+        }
+      },
+    });
+
+    expect(calls).toEqual([
+      ["one", "two", "three", "four"],
+      ["one", "two"],
+      ["three", "four"],
+    ]);
+    expect(onSplitGroup).toHaveBeenCalledWith(
+      expect.objectContaining({
+        error: uploadTooLarge,
+        group: ["one", "two", "three", "four"],
+        parts: [
+          ["one", "two"],
+          ["three", "four"],
+        ],
+        depth: 0,
+      }),
+    );
+  });
+
+  it("does not split a single rejected batch request", async () => {
+    const uploadTooLarge = new Error("batch upload failed: 413 payload too large");
+
+    await expect(
+      runEmbeddingBatchGroups({
+        requests: ["one"],
+        maxRequests: 100,
+        wait: true,
+        pollIntervalMs: 1000,
+        timeoutMs: 60_000,
+        concurrency: 1,
+        debugLabel: "embedding batch submit",
+        shouldSplitGroupOnError: () => true,
+        runGroup: async () => {
+          throw uploadTooLarge;
+        },
+      }),
+    ).rejects.toThrow(uploadTooLarge);
+  });
+});

packages/memory-host-sdk/src/host/batch-runner.ts

@@ -0,0 +1,141 @@
+// Memory Host SDK module implements batch runner behavior.
+import { resolveSafeTimeoutDelayMs } from "../../../gateway-client/src/timeouts.js";
+import { splitBatchRequestsByLimits } from "./batch-utils.js";
+import { runWithConcurrency } from "./internal.js";
+
+// Shared runner for splitting and executing remote embedding batch groups.
+
+/** Execution controls for provider embedding batch submissions and polling. */
+export type EmbeddingBatchExecutionParams = {
+  wait: boolean;
+  pollIntervalMs: number;
+  timeoutMs: number;
+  concurrency: number;
+  debug?: (message: string, data?: Record<string, unknown>) => void;
+};
+
+type EmbeddingBatchGroupRunArgs<TRequest> = {
+  group: TRequest[];
+  groupIndex: number;
+  groups: number;
+  byCustomId: Map<string, number[]>;
+  pollIntervalMs: number;
+  timeoutMs: number;
+};
+
+type EmbeddingBatchSplitArgs<TRequest> = {
+  error: unknown;
+  group: TRequest[];
+  parts: TRequest[][];
+  groupIndex: number;
+  groups: number;
+  depth: number;
+};
+
+/** Clamp polling to both configured poll interval and total timeout budget. */
+function resolveEmbeddingBatchPollIntervalMs(params: {
+  pollIntervalMs: number;
+  timeoutMs: number;
+}): number {
+  const safePollIntervalMs = resolveSafeTimeoutDelayMs(params.pollIntervalMs);
+  const safeTimeoutMs =
+    typeof params.timeoutMs === "number" &&
+    Number.isFinite(params.timeoutMs) &&
+    params.timeoutMs > 0
+      ? resolveSafeTimeoutDelayMs(params.timeoutMs)
+      : safePollIntervalMs;
+  return Math.min(safePollIntervalMs, safeTimeoutMs);
+}
+
+/** Run request groups with bounded concurrency and return embeddings by custom id. */
+export async function runEmbeddingBatchGroups<TRequest>(params: {
+  requests: TRequest[];
+  maxRequests: number;
+  maxJsonlBytes?: number;
+  wait: EmbeddingBatchExecutionParams["wait"];
+  pollIntervalMs: EmbeddingBatchExecutionParams["pollIntervalMs"];
+  timeoutMs: EmbeddingBatchExecutionParams["timeoutMs"];
+  concurrency: EmbeddingBatchExecutionParams["concurrency"];
+  debugLabel: string;
+  debug?: EmbeddingBatchExecutionParams["debug"];
+  shouldSplitGroupOnError?: (error: unknown, group: TRequest[]) => boolean;
+  onSplitGroup?: (args: EmbeddingBatchSplitArgs<TRequest>) => void;
+  runGroup: (args: EmbeddingBatchGroupRunArgs<TRequest>) => Promise<void>;
+}): Promise<Map<string, number[]>> {
+  if (params.requests.length === 0) {
+    return new Map();
+  }
+  const groups = splitBatchRequestsByLimits(params.requests, {
+    maxRequests: params.maxRequests,
+    maxJsonlBytes: params.maxJsonlBytes,
+  });
+  const byCustomId = new Map<string, number[]>();
+  const pollIntervalMs = resolveEmbeddingBatchPollIntervalMs(params);
+  const runGroup = async (group: TRequest[], groupIndex: number, depth = 0): Promise<void> => {
+    try {
+      await params.runGroup({
+        group,
+        groupIndex,
+        groups: groups.length,
+        byCustomId,
+        pollIntervalMs,
+        timeoutMs: params.timeoutMs,
+      });
+    } catch (error) {
+      if (group.length <= 1 || !params.shouldSplitGroupOnError?.(error, group)) {
+        throw error;
+      }
+      const splitAt = Math.ceil(group.length / 2);
+      const parts = [group.slice(0, splitAt), group.slice(splitAt)].filter(
+        (part) => part.length > 0,
+      );
+      params.onSplitGroup?.({
+        error,
+        group,
+        parts,
+        groupIndex,
+        groups: groups.length,
+        depth,
+      });
+      for (const part of parts) {
+        await runGroup(part, groupIndex, depth + 1);
+      }
+    }
+  };
+  const tasks = groups.map((group, groupIndex) => async () => {
+    await runGroup(group, groupIndex);
+  });
+
+  params.debug?.(params.debugLabel, {
+    requests: params.requests.length,
+    groups: groups.length,
+    maxRequests: params.maxRequests,
+    maxJsonlBytes: params.maxJsonlBytes,
+    wait: params.wait,
+    concurrency: params.concurrency,
+    pollIntervalMs,
+    timeoutMs: params.timeoutMs,
+  });
+
+  await runWithConcurrency(tasks, params.concurrency);
+  return byCustomId;
+}
+
+/** Build normalized batch-group options for provider-specific runners. */
+export function buildEmbeddingBatchGroupOptions<TRequest>(
+  params: { requests: TRequest[] } & EmbeddingBatchExecutionParams,
+  options: { maxRequests: number; maxJsonlBytes?: number; debugLabel: string },
+) {
+  const pollIntervalMs = resolveEmbeddingBatchPollIntervalMs(params);
+  return {
+    requests: params.requests,
+    maxRequests: options.maxRequests,
+    maxJsonlBytes: options.maxJsonlBytes,
+    wait: params.wait,
+    pollIntervalMs,
+    timeoutMs: params.timeoutMs,
+    concurrency: params.concurrency,
+    debug: params.debug,
+    debugLabel: options.debugLabel,
+  };
+}

packages/memory-host-sdk/src/host/batch-status.test.ts

@@ -0,0 +1,61 @@
+// Memory Host SDK tests cover batch status behavior.
+import { describe, expect, it } from "vitest";
+import {
+  resolveBatchCompletionFromStatus,
+  resolveCompletedBatchResult,
+  throwIfBatchTerminalFailure,
+} from "./batch-status.js";
+
+describe("batch-status helpers", () => {
+  it("resolves completion payload from completed status", () => {
+    expect(
+      resolveBatchCompletionFromStatus({
+        provider: "openai",
+        batchId: "b1",
+        status: {
+          output_file_id: "out-1",
+          error_file_id: "err-1",
+        },
+      }),
+    ).toEqual({
+      outputFileId: "out-1",
+      errorFileId: "err-1",
+    });
+  });
+
+  it("throws for terminal failure states", async () => {
+    await expect(
+      throwIfBatchTerminalFailure({
+        provider: "voyage",
+        status: { id: "b2", status: "failed", error_file_id: "err-file" },
+        readError: async () => "bad input",
+      }),
+    ).rejects.toThrow("voyage batch b2 failed: bad input");
+  });
+
+  it("returns completed result directly without waiting", async () => {
+    const waitForBatch = async () => ({ outputFileId: "out-2" });
+    const result = await resolveCompletedBatchResult({
+      provider: "openai",
+      status: {
+        id: "b3",
+        status: "completed",
+        output_file_id: "out-3",
+      },
+      wait: false,
+      waitForBatch,
+    });
+    expect(result).toEqual({ outputFileId: "out-3", errorFileId: undefined });
+  });
+
+  it("throws when wait disabled and batch is not complete", async () => {
+    await expect(
+      resolveCompletedBatchResult({
+        provider: "openai",
+        status: { id: "b4", status: "pending" },
+        wait: false,
+        waitForBatch: async () => ({ outputFileId: "out" }),
+      }),
+    ).rejects.toThrow("openai batch b4 submitted; enable remote.batch.wait to await completion");
+  });
+});

packages/memory-host-sdk/src/host/batch-status.ts

@@ -0,0 +1,76 @@
+// Batch status helpers shared by remote embedding providers.
+
+const TERMINAL_FAILURE_STATES = new Set(["failed", "expired", "cancelled", "canceled"]);
+
+/** Minimal provider batch status used for completion and terminal-failure checks. */
+type BatchStatusLike = {
+  id?: string;
+  status?: string;
+  output_file_id?: string | null;
+  error_file_id?: string | null;
+};
+
+/** File ids returned once a batch has completed. */
+export type BatchCompletionResult = {
+  outputFileId: string;
+  errorFileId?: string;
+};
+
+/** Convert a completed provider status payload into output/error file ids. */
+export function resolveBatchCompletionFromStatus(params: {
+  provider: string;
+  batchId: string;
+  status: BatchStatusLike;
+}): BatchCompletionResult {
+  if (!params.status.output_file_id) {
+    throw new Error(`${params.provider} batch ${params.batchId} completed without output file`);
+  }
+  return {
+    outputFileId: params.status.output_file_id,
+    errorFileId: params.status.error_file_id ?? undefined,
+  };
+}
+
+/** Throw when a provider reports a terminal failure, including error-file detail if available. */
+export async function throwIfBatchTerminalFailure(params: {
+  provider: string;
+  status: BatchStatusLike;
+  readError: (errorFileId: string) => Promise<string | undefined>;
+}): Promise<void> {
+  const state = params.status.status ?? "unknown";
+  if (!TERMINAL_FAILURE_STATES.has(state)) {
+    return;
+  }
+  const detail = params.status.error_file_id
+    ? await params.readError(params.status.error_file_id)
+    : undefined;
+  const suffix = detail ? `: ${detail}` : "";
+  throw new Error(`${params.provider} batch ${params.status.id ?? "<unknown>"} ${state}${suffix}`);
+}
+
+/** Resolve the completed batch files, optionally waiting according to caller policy. */
+export async function resolveCompletedBatchResult(params: {
+  provider: string;
+  status: BatchStatusLike;
+  wait: boolean;
+  waitForBatch: () => Promise<BatchCompletionResult>;
+}): Promise<BatchCompletionResult> {
+  const batchId = params.status.id ?? "<unknown>";
+  if (!params.wait && params.status.status !== "completed") {
+    throw new Error(
+      `${params.provider} batch ${batchId} submitted; enable remote.batch.wait to await completion`,
+    );
+  }
+  const completed =
+    params.status.status === "completed"
+      ? resolveBatchCompletionFromStatus({
+          provider: params.provider,
+          batchId,
+          status: params.status,
+        })
+      : await params.waitForBatch();
+  if (!completed.outputFileId) {
+    throw new Error(`${params.provider} batch ${batchId} completed without output file`);
+  }
+  return completed;
+}

packages/memory-host-sdk/src/host/batch-upload.test.ts

@@ -0,0 +1,194 @@
+// Memory Host SDK tests cover batch upload behavior.
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import { uploadBatchJsonlFile } from "./batch-upload.js";
+import { withRemoteHttpResponse } from "./remote-http.js";
+
+vi.mock("./remote-http.js", () => ({
+  withRemoteHttpResponse: vi.fn(),
+}));
+
+const remoteHttpMock = vi.mocked(withRemoteHttpResponse);
+
+function textResponse(body: string, status: number): Response {
+  return new Response(body, { status });
+}
+
+function streamingTextResponse(params: {
+  body: string;
+  status: number;
+  headers?: HeadersInit;
+  onCancel: () => void;
+}): Response {
+  const encoded = new TextEncoder().encode(params.body);
+  const stream = new ReadableStream<Uint8Array>({
+    start(controller) {
+      controller.enqueue(encoded);
+    },
+    cancel() {
+      params.onCancel();
+    },
+  });
+  return new Response(stream, { status: params.status, headers: params.headers });
+}
+
+function stallingResponse(params: { status: number; onCancel: () => void }): Response {
+  const reader = {
+    read: () => new Promise<ReadableStreamReadResult<Uint8Array>>(() => {}),
+    cancel: async () => {
+      params.onCancel();
+    },
+    releaseLock: () => undefined,
+  } as ReadableStreamDefaultReader<Uint8Array>;
+
+  return {
+    status: params.status,
+    ok: params.status >= 200 && params.status < 300,
+    headers: new Headers(),
+    body: { getReader: () => reader },
+  } as Response;
+}
+
+describe("uploadBatchJsonlFile", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+  });
+
+  it("wraps malformed file-upload JSON with the request error prefix", async () => {
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(textResponse("{ nope", 200));
+    });
+
+    await expect(
+      uploadBatchJsonlFile({
+        client: {
+          baseUrl: "https://memory.example/v1",
+          headers: { Authorization: "Bearer test" },
+        },
+        requests: [{ input: "one" }],
+        errorPrefix: "file upload failed",
+      }),
+    ).rejects.toThrow("file upload failed: malformed JSON response");
+  });
+
+  it("bounds non-ok file-upload response bodies before formatting the error", async () => {
+    let canceled = false;
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(
+        streamingTextResponse({
+          body: "x".repeat(12_000),
+          status: 413,
+          onCancel: () => {
+            canceled = true;
+          },
+        }),
+      );
+    });
+
+    await expect(
+      uploadBatchJsonlFile({
+        client: {
+          baseUrl: "https://memory.example/v1",
+          headers: { Authorization: "Bearer test" },
+        },
+        requests: [{ input: "one" }],
+        errorPrefix: "file upload failed",
+      }),
+    ).rejects.toThrow(`file upload failed: 413 ${"x".repeat(1_000)}... [truncated]`);
+    expect(canceled).toBe(true);
+  });
+
+  it("rejects oversized successful file-upload JSON before parsing", async () => {
+    let canceled = false;
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(
+        streamingTextResponse({
+          body: '{"id":"file_123"}',
+          status: 200,
+          headers: { "content-length": "64" },
+          onCancel: () => {
+            canceled = true;
+          },
+        }),
+      );
+    });
+
+    await expect(
+      uploadBatchJsonlFile({
+        client: {
+          baseUrl: "https://memory.example/v1",
+          headers: { Authorization: "Bearer test" },
+        },
+        requests: [{ input: "one" }],
+        errorPrefix: "file upload failed",
+        maxResponseBytes: 8,
+      }),
+    ).rejects.toThrow("file upload failed: response body too large: 64 bytes (limit: 8 bytes)");
+    expect(canceled).toBe(true);
+  });
+
+  it("passes caller abort signals through non-ok file-upload response snippets", async () => {
+    let canceled = false;
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(
+        stallingResponse({
+          status: 500,
+          onCancel: () => {
+            canceled = true;
+          },
+        }),
+      );
+    });
+    const controller = new AbortController();
+    const upload = uploadBatchJsonlFile({
+      client: {
+        baseUrl: "https://memory.example/v1",
+        headers: { Authorization: "Bearer test" },
+      },
+      requests: [{ input: "one" }],
+      errorPrefix: "file upload failed",
+      signal: controller.signal,
+    });
+
+    await new Promise((resolve) => {
+      setTimeout(resolve, 0);
+    });
+    controller.abort(new Error("upload aborted"));
+
+    await expect(upload).rejects.toThrow("upload aborted");
+    expect(canceled).toBe(true);
+    expect(remoteHttpMock.mock.calls[0]?.[0].signal).toBe(controller.signal);
+  });
+
+  it("passes caller abort signals through successful file-upload JSON reads", async () => {
+    let canceled = false;
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(
+        stallingResponse({
+          status: 200,
+          onCancel: () => {
+            canceled = true;
+          },
+        }),
+      );
+    });
+    const controller = new AbortController();
+    const upload = uploadBatchJsonlFile({
+      client: {
+        baseUrl: "https://memory.example/v1",
+        headers: { Authorization: "Bearer test" },
+      },
+      requests: [{ input: "one" }],
+      errorPrefix: "file upload failed",
+      signal: controller.signal,
+    });
+
+    await new Promise((resolve) => {
+      setTimeout(resolve, 0);
+    });
+    controller.abort(new Error("upload json aborted"));
+
+    await expect(upload).rejects.toThrow("upload json aborted");
+    expect(canceled).toBe(true);
+    expect(remoteHttpMock.mock.calls[0]?.[0].signal).toBe(controller.signal);
+  });
+});

packages/memory-host-sdk/src/host/batch-upload.ts

@@ -0,0 +1,57 @@
+// Memory Host SDK module implements batch upload behavior.
+import {
+  buildBatchHeaders,
+  normalizeBatchBaseUrl,
+  type BatchHttpClientConfig,
+} from "./batch-utils.js";
+import { hashText } from "./hash.js";
+import { withRemoteHttpResponse } from "./remote-http.js";
+import { readResponseJsonWithLimit, readResponseTextSnippet } from "./response-snippet.js";
+
+// Uploads provider batch JSONL payloads through the shared remote HTTP guard.
+
+/** Upload embedding batch requests and return the provider file id. */
+export async function uploadBatchJsonlFile(params: {
+  client: BatchHttpClientConfig;
+  requests: unknown[];
+  errorPrefix: string;
+  maxResponseBytes?: number;
+  signal?: AbortSignal;
+}): Promise<string> {
+  const baseUrl = normalizeBatchBaseUrl(params.client);
+  const jsonl = params.requests.map((request) => JSON.stringify(request)).join("\n");
+  const form = new FormData();
+  form.append("purpose", "batch");
+  form.append(
+    "file",
+    new Blob([jsonl], { type: "application/jsonl" }),
+    `memory-embeddings.${hashText(String(Date.now()))}.jsonl`,
+  );
+
+  const filePayload = await withRemoteHttpResponse({
+    url: `${baseUrl}/files`,
+    ssrfPolicy: params.client.ssrfPolicy,
+    fetchImpl: params.client.fetchImpl,
+    signal: params.signal,
+    init: {
+      method: "POST",
+      headers: buildBatchHeaders(params.client, { json: false }),
+      body: form,
+    },
+    onResponse: async (fileRes) => {
+      if (!fileRes.ok) {
+        const text = await readResponseTextSnippet(fileRes, { signal: params.signal });
+        throw new Error(`${params.errorPrefix}: ${fileRes.status} ${text}`);
+      }
+      return (await readResponseJsonWithLimit(fileRes, {
+        errorPrefix: params.errorPrefix,
+        maxBytes: params.maxResponseBytes,
+        signal: params.signal,
+      })) as { id?: string };
+    },
+  });
+  if (!filePayload.id) {
+    throw new Error(`${params.errorPrefix}: missing file id`);
+  }
+  return filePayload.id;
+}

packages/memory-host-sdk/src/host/batch-utils.ts

@@ -0,0 +1,94 @@
+// Memory Host SDK helper module supports batch utils behavior.
+import type { SsrFPolicy } from "./ssrf-policy.js";
+
+// Common HTTP and grouping helpers for remote embedding batch clients.
+
+/** Minimal HTTP client config needed by batch providers. */
+export type BatchHttpClientConfig = {
+  baseUrl?: string;
+  headers?: Record<string, string>;
+  ssrfPolicy?: SsrFPolicy;
+  fetchImpl?: typeof fetch;
+};
+
+/** Normalize batch API base URLs by removing one trailing slash. */
+export function normalizeBatchBaseUrl(client: BatchHttpClientConfig): string {
+  return client.baseUrl?.replace(/\/$/, "") ?? "";
+}
+
+/** Build request headers, preserving caller auth and controlling JSON/form content type. */
+export function buildBatchHeaders(
+  client: Pick<BatchHttpClientConfig, "headers">,
+  params: { json: boolean },
+): Record<string, string> {
+  const headers = client.headers ? { ...client.headers } : {};
+  if (params.json) {
+    if (!headers["Content-Type"] && !headers["content-type"]) {
+      headers["Content-Type"] = "application/json";
+    }
+  } else {
+    delete headers["Content-Type"];
+    delete headers["content-type"];
+  }
+  return headers;
+}
+
+const jsonlEncoder = new TextEncoder();
+
+function estimateJsonlLineBytes(request: unknown): number {
+  return jsonlEncoder.encode(JSON.stringify(request) ?? "").byteLength;
+}
+
+function normalizePositiveInteger(value: number | undefined): number | undefined {
+  if (typeof value !== "number" || !Number.isFinite(value) || value <= 0) {
+    return undefined;
+  }
+  return Math.floor(value);
+}
+
+/** Split provider requests into max-sized groups while preserving order. */
+export function splitBatchRequests<T>(requests: T[], maxRequests: number): T[][] {
+  const limit = normalizePositiveInteger(maxRequests) ?? 1;
+  if (requests.length <= limit) {
+    return [requests];
+  }
+  const groups: T[][] = [];
+  for (let i = 0; i < requests.length; i += limit) {
+    groups.push(requests.slice(i, i + limit));
+  }
+  return groups;
+}
+
+export function splitBatchRequestsByLimits<T>(
+  requests: T[],
+  limits: { maxRequests: number; maxJsonlBytes?: number },
+): T[][] {
+  const maxRequests = normalizePositiveInteger(limits.maxRequests) ?? 1;
+  const maxJsonlBytes = normalizePositiveInteger(limits.maxJsonlBytes);
+  if (!maxJsonlBytes) {
+    return splitBatchRequests(requests, maxRequests);
+  }
+
+  const groups: T[][] = [];
+  let current: T[] = [];
+  let currentBytes = 0;
+  for (const request of requests) {
+    const requestBytes = estimateJsonlLineBytes(request);
+    const separatorBytes = current.length === 0 ? 0 : 1;
+    const wouldExceedRequests = current.length >= maxRequests;
+    const wouldExceedBytes =
+      current.length > 0 && currentBytes + separatorBytes + requestBytes > maxJsonlBytes;
+    if (current.length > 0 && (wouldExceedRequests || wouldExceedBytes)) {
+      groups.push(current);
+      current = [];
+      currentBytes = 0;
+    }
+
+    currentBytes += (current.length === 0 ? 0 : 1) + requestBytes;
+    current.push(request);
+  }
+  if (current.length > 0) {
+    groups.push(current);
+  }
+  return groups;
+}

packages/memory-host-sdk/src/host/config-utils.test.ts

@@ -0,0 +1,15 @@
+// Memory Host SDK tests cover config utils behavior.
+import { describe, expect, it } from "vitest";
+import { parseDurationMs } from "./config-utils.js";
+
+describe("parseDurationMs", () => {
+  it("parses decimal durations into milliseconds", () => {
+    expect(parseDurationMs("1.5s")).toBe(1_500);
+    expect(parseDurationMs("1h30m")).toBe(5_400_000);
+  });
+
+  it("rejects unsafe millisecond results", () => {
+    expect(() => parseDurationMs("9007199254740993ms")).toThrow(/invalid duration/u);
+    expect(() => parseDurationMs("9007199254740990ms10ms")).toThrow(/invalid duration/u);
+  });
+});

packages/memory-host-sdk/src/host/config-utils.ts

@@ -0,0 +1,423 @@
+// Memory Host SDK helper module supports config utils behavior.
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+import {
+  normalizeLowercaseStringOrEmpty,
+  normalizeOptionalString,
+  normalizeStringEntries,
+  uniqueStrings,
+} from "./string-utils.js";
+export { splitShellArgs } from "./openclaw-runtime-io.js";
+
+// Shared OpenClaw config helpers used by memory host, QMD, and agent context code.
+
+/** Chat shape used by memory send-policy matching. */
+export type ChatType = "direct" | "group" | "channel";
+/** Memory backend selected by user config. */
+export type MemoryBackend = "builtin" | "qmd";
+/** Citation injection behavior for memory search results. */
+export type MemoryCitationsMode = "auto" | "on" | "off";
+/** QMD command mode used for search calls. */
+export type MemoryQmdSearchMode = "query" | "search" | "vsearch";
+/** QMD startup policy for background indexing. */
+export type MemoryQmdStartupMode = "off" | "idle" | "immediate";
+
+/** Action returned by a session send-policy rule. */
+export type SessionSendPolicyAction = "allow" | "deny";
+/** Match criteria for one memory send-policy rule. */
+export type SessionSendPolicyMatch = {
+  channel?: string;
+  chatType?: ChatType;
+  keyPrefix?: string;
+  rawKeyPrefix?: string;
+};
+/** One ordered rule in session send-policy config. */
+export type SessionSendPolicyRule = {
+  action: SessionSendPolicyAction;
+  match?: SessionSendPolicyMatch;
+};
+/** Memory send-policy config with default action and ordered rules. */
+export type SessionSendPolicyConfig = {
+  default?: SessionSendPolicyAction;
+  rules?: SessionSendPolicyRule[];
+};
+
+/** QMD collection path plus optional display name and glob pattern. */
+export type MemoryQmdIndexPath = {
+  path: string;
+  name?: string;
+  pattern?: string;
+};
+
+/** QMD mcporter daemon integration config. */
+export type MemoryQmdMcporterConfig = {
+  enabled?: boolean;
+  serverName?: string;
+  startDaemon?: boolean;
+};
+
+/** QMD session export config. */
+export type MemoryQmdSessionConfig = {
+  enabled?: boolean;
+  exportDir?: string;
+  retentionDays?: number;
+};
+
+/** QMD update, debounce, startup, and timeout config. */
+export type MemoryQmdUpdateConfig = {
+  interval?: string;
+  debounceMs?: number;
+  onBoot?: boolean;
+  startup?: MemoryQmdStartupMode;
+  startupDelayMs?: number;
+  waitForBootSync?: boolean;
+  embedInterval?: string;
+  commandTimeoutMs?: number;
+  updateTimeoutMs?: number;
+  embedTimeoutMs?: number;
+};
+
+/** Search and injection limits for QMD memory results. */
+export type MemoryQmdLimitsConfig = {
+  maxResults?: number;
+  maxSnippetChars?: number;
+  maxInjectedChars?: number;
+  timeoutMs?: number;
+};
+
+/** Full QMD-backed memory config. */
+export type MemoryQmdConfig = {
+  command?: string;
+  mcporter?: MemoryQmdMcporterConfig;
+  searchMode?: MemoryQmdSearchMode;
+  rerank?: boolean;
+  searchTool?: string;
+  includeDefaultMemory?: boolean;
+  paths?: MemoryQmdIndexPath[];
+  sessions?: MemoryQmdSessionConfig;
+  update?: MemoryQmdUpdateConfig;
+  limits?: MemoryQmdLimitsConfig;
+  scope?: SessionSendPolicyConfig;
+};
+
+/** Top-level memory config shared by host and runtime callers. */
+export type MemoryConfig = {
+  backend?: MemoryBackend;
+  citations?: MemoryCitationsMode;
+  qmd?: MemoryQmdConfig;
+};
+
+/** Per-agent memory search enablement and extra collection paths. */
+export type MemorySearchConfig = {
+  enabled?: boolean;
+  extraPaths?: string[];
+  qmd?: {
+    extraCollections?: MemoryQmdIndexPath[];
+  };
+};
+
+/** Agent context limits that bound memory file reads. */
+export type AgentContextLimitsConfig = {
+  memoryGetMaxChars?: number;
+  memoryGetDefaultLines?: number;
+};
+
+/** Secret reference accepted by provider header config. */
+export type SecretInput =
+  | string
+  | {
+      source: string;
+      provider: string;
+      id: string;
+    };
+
+/** Agent-level config fields consumed by memory host helpers. */
+type AgentConfig = {
+  id?: string;
+  default?: boolean;
+  workspace?: string;
+  memorySearch?: MemorySearchConfig;
+  contextLimits?: AgentContextLimitsConfig;
+};
+
+/** Narrow OpenClaw config shape consumed by memory host utilities. */
+export type OpenClawConfig = {
+  agents?: {
+    defaults?: {
+      workspace?: string;
+      memorySearch?: MemorySearchConfig;
+      contextLimits?: AgentContextLimitsConfig;
+    };
+    list?: AgentConfig[];
+  };
+  memory?: MemoryConfig;
+  models?: {
+    providers?: Record<
+      string,
+      {
+        api?: string;
+        baseUrl?: string;
+        headers?: Record<string, SecretInput>;
+      }
+    >;
+  };
+};
+
+/** Root memory filename used in agent workspaces. */
+export const CANONICAL_ROOT_MEMORY_FILENAME = "MEMORY.md";
+
+const DEFAULT_AGENT_ID = "main";
+const VALID_ID_RE = /^[a-z0-9][a-z0-9_-]{0,63}$/i;
+const INVALID_CHARS_RE = /[^a-z0-9_-]+/g;
+const LEADING_DASH_RE = /^-+/;
+const TRAILING_DASH_RE = /-+$/;
+const LEGACY_STATE_DIRNAMES = [".clawdbot"] as const;
+const NEW_STATE_DIRNAME = ".openclaw";
+const DURATION_MULTIPLIERS: Record<string, number> = {
+  ms: 1,
+  s: 1000,
+  m: 60_000,
+  h: 3_600_000,
+  d: 86_400_000,
+};
+
+/** Round parsed durations and reject values outside the safe integer range. */
+function roundDurationMs(raw: string, value: number): number {
+  const rounded = Math.round(value);
+  if (!Number.isSafeInteger(rounded)) {
+    throw new Error(`invalid duration: ${raw}`);
+  }
+  return rounded;
+}
+
+/** Normalize user or config agent ids to the filesystem-safe canonical form. */
+export function normalizeAgentId(value: string | undefined | null): string {
+  const trimmed = (value ?? "").trim();
+  if (!trimmed) {
+    return DEFAULT_AGENT_ID;
+  }
+  const normalized = normalizeLowercaseStringOrEmpty(trimmed);
+  if (VALID_ID_RE.test(trimmed)) {
+    return normalized;
+  }
+  return (
+    normalized
+      .replace(INVALID_CHARS_RE, "-")
+      .replace(LEADING_DASH_RE, "")
+      .replace(TRAILING_DASH_RE, "")
+      .slice(0, 64) || DEFAULT_AGENT_ID
+  );
+}
+
+/** Treat shell-placeholder home values as absent. */
+function normalizeHomeValue(value: string | undefined): string | undefined {
+  const trimmed = normalizeOptionalString(value);
+  if (!trimmed || trimmed === "undefined" || trimmed === "null") {
+    return undefined;
+  }
+  return trimmed;
+}
+
+/** Resolve the underlying OS home before applying OpenClaw-specific overrides. */
+function resolveRawOsHomeDir(env: NodeJS.ProcessEnv, homedir: () => string): string | undefined {
+  return (
+    normalizeHomeValue(env.HOME) ??
+    normalizeHomeValue(env.USERPROFILE) ??
+    normalizeHomeValue(homedir())
+  );
+}
+
+/** Resolve OPENCLAW_HOME or the OS home, falling back to cwd for hermetic tests. */
+function resolveRequiredHomeDir(
+  env: NodeJS.ProcessEnv = process.env,
+  homedir: () => string = os.homedir,
+): string {
+  const explicitHome = normalizeHomeValue(env.OPENCLAW_HOME);
+  const rawHome = explicitHome
+    ? explicitHome.replace(/^~(?=$|[\\/])/, resolveRawOsHomeDir(env, homedir) ?? "")
+    : resolveRawOsHomeDir(env, homedir);
+  return rawHome ? path.resolve(rawHome) : path.resolve(process.cwd());
+}
+
+/** Resolve absolute user paths, including "~" against the effective OpenClaw home. */
+export function resolveUserPath(
+  input: string,
+  env: NodeJS.ProcessEnv = process.env,
+  homedir: () => string = os.homedir,
+): string {
+  const trimmed = input.trim();
+  if (!trimmed) {
+    return trimmed;
+  }
+  if (trimmed.startsWith("~")) {
+    return path.resolve(trimmed.replace(/^~(?=$|[\\/])/, resolveRequiredHomeDir(env, homedir)));
+  }
+  return path.resolve(trimmed);
+}
+
+/** Return legacy state roots in priority order. */
+function legacyStateDirs(homedir: () => string): string[] {
+  return LEGACY_STATE_DIRNAMES.map((dir) => path.join(homedir(), dir));
+}
+
+/** Resolve the current state root while preserving shipped legacy installs when present. */
+export function resolveStateDir(
+  env: NodeJS.ProcessEnv = process.env,
+  homedir: () => string = os.homedir,
+): string {
+  const override = env.OPENCLAW_STATE_DIR?.trim();
+  if (override) {
+    return resolveUserPath(override, env, homedir);
+  }
+  const effectiveHome = () => resolveRequiredHomeDir(env, homedir);
+  const nextDir = path.join(effectiveHome(), NEW_STATE_DIRNAME);
+  if (env.OPENCLAW_TEST_FAST === "1" || fs.existsSync(nextDir)) {
+    return nextDir;
+  }
+  // Existing legacy state remains authoritative until an explicit migration creates .openclaw.
+  const existingLegacy = legacyStateDirs(effectiveHome).find((dir) => {
+    try {
+      return fs.existsSync(dir);
+    } catch {
+      return false;
+    }
+  });
+  return existingLegacy ?? nextDir;
+}
+
+/** Resolve the default agent workspace, partitioned by OPENCLAW_PROFILE when set. */
+function resolveDefaultAgentWorkspaceDir(env: NodeJS.ProcessEnv = process.env): string {
+  const home = resolveRequiredHomeDir(env, os.homedir);
+  const profile = env.OPENCLAW_PROFILE?.trim();
+  if (profile && normalizeLowercaseStringOrEmpty(profile) !== "default") {
+    return path.join(home, ".openclaw", `workspace-${profile}`);
+  }
+  return path.join(home, ".openclaw", "workspace");
+}
+
+/** Return configured agent entries after dropping nullish placeholders. */
+function listAgentEntries(cfg: OpenClawConfig): AgentConfig[] {
+  return Array.isArray(cfg.agents?.list)
+    ? cfg.agents.list.filter((entry): entry is AgentConfig => Boolean(entry))
+    : [];
+}
+
+/** Resolve the default agent id from explicit default marker or first agent entry. */
+function resolveDefaultAgentId(cfg: OpenClawConfig): string {
+  const agents = listAgentEntries(cfg);
+  if (agents.length === 0) {
+    return DEFAULT_AGENT_ID;
+  }
+  const chosen = (agents.find((agent) => agent.default) ?? agents[0])?.id;
+  return normalizeAgentId(chosen || DEFAULT_AGENT_ID);
+}
+
+/** Find one agent config by canonical id. */
+function resolveAgentConfig(cfg: OpenClawConfig, agentId: string): AgentConfig | undefined {
+  const id = normalizeAgentId(agentId);
+  return listAgentEntries(cfg).find((entry) => normalizeAgentId(entry.id) === id);
+}
+
+/** Remove null bytes before paths are handed to filesystem APIs. */
+function stripNullBytes(value: string): string {
+  return value.replaceAll("\0", "");
+}
+
+/** Resolve the workspace directory for an agent id and config defaults. */
+export function resolveAgentWorkspaceDir(
+  cfg: OpenClawConfig,
+  agentId: string,
+  env: NodeJS.ProcessEnv = process.env,
+): string {
+  const id = normalizeAgentId(agentId);
+  const configured = resolveAgentConfig(cfg, id)?.workspace?.trim();
+  if (configured) {
+    return stripNullBytes(resolveUserPath(configured, env));
+  }
+  const fallback = cfg.agents?.defaults?.workspace?.trim();
+  if (id === resolveDefaultAgentId(cfg)) {
+    return stripNullBytes(
+      fallback ? resolveUserPath(fallback, env) : resolveDefaultAgentWorkspaceDir(env),
+    );
+  }
+  if (fallback) {
+    return stripNullBytes(path.join(resolveUserPath(fallback, env), id));
+  }
+  return stripNullBytes(path.join(resolveStateDir(env), `workspace-${id}`));
+}
+
+/** Resolve context limits for an agent with defaults fallback. */
+export function resolveAgentContextLimits(
+  cfg: OpenClawConfig | undefined,
+  agentId?: string | null,
+): AgentContextLimitsConfig | undefined {
+  const defaults = cfg?.agents?.defaults?.contextLimits;
+  if (!cfg || !agentId) {
+    return defaults;
+  }
+  return resolveAgentConfig(cfg, agentId)?.contextLimits ?? defaults;
+}
+
+/** Resolve enabled memory search config plus deduplicated extra paths for an agent. */
+export function resolveMemorySearchConfig(
+  cfg: OpenClawConfig,
+  agentId: string,
+): { enabled: boolean; extraPaths: string[] } | null {
+  const defaults = cfg.agents?.defaults?.memorySearch;
+  const overrides = resolveAgentConfig(cfg, agentId)?.memorySearch;
+  const enabled = overrides?.enabled ?? defaults?.enabled ?? true;
+  if (!enabled) {
+    return null;
+  }
+  const rawPaths = normalizeStringEntries([
+    ...(defaults?.extraPaths ?? []),
+    ...(overrides?.extraPaths ?? []),
+  ]);
+  return {
+    enabled,
+    extraPaths: uniqueStrings(rawPaths),
+  };
+}
+
+/** Parse compact duration strings such as "500ms", "5s", or "1h30m" into milliseconds. */
+export function parseDurationMs(
+  raw: string,
+  opts?: { defaultUnit?: "ms" | "s" | "m" | "h" | "d" },
+): number {
+  const trimmed = normalizeLowercaseStringOrEmpty(normalizeOptionalString(raw) ?? "");
+  if (!trimmed) {
+    throw new Error("invalid duration (empty)");
+  }
+  const single = /^(\d+(?:\.\d+)?)(ms|s|m|h|d)?$/.exec(trimmed);
+  if (single) {
+    const value = Number(single[1]);
+    if (!Number.isFinite(value) || value < 0) {
+      throw new Error(`invalid duration: ${raw}`);
+    }
+    const unit = single[2] ?? opts?.defaultUnit ?? "ms";
+    return roundDurationMs(raw, value * (DURATION_MULTIPLIERS[unit] ?? 1));
+  }
+
+  let totalMs = 0;
+  let consumed = 0;
+  const tokenRe = /(\d+(?:\.\d+)?)(ms|s|m|h|d)/g;
+  for (const match of trimmed.matchAll(tokenRe)) {
+    const [full, valueRaw, unitRaw] = match;
+    const index = match.index ?? -1;
+    if (!full || !valueRaw || !unitRaw || index !== consumed) {
+      throw new Error(`invalid duration: ${raw}`);
+    }
+    const value = Number(valueRaw);
+    const multiplier = DURATION_MULTIPLIERS[unitRaw];
+    if (!Number.isFinite(value) || value < 0 || !multiplier) {
+      throw new Error(`invalid duration: ${raw}`);
+    }
+    totalMs += value * multiplier;
+    consumed += full.length;
+  }
+  if (consumed !== trimmed.length || consumed === 0) {
+    throw new Error(`invalid duration: ${raw}`);
+  }
+  return roundDurationMs(raw, totalMs);
+}

packages/memory-host-sdk/src/host/embedding-chunk-limits.test.ts

@@ -0,0 +1,142 @@
+// Memory Host SDK tests cover embedding chunk limits behavior.
+import { describe, expect, it } from "vitest";
+import { enforceEmbeddingMaxInputTokens } from "./embedding-chunk-limits.js";
+import { estimateUtf8Bytes } from "./embedding-input-limits.js";
+import type { EmbeddingProvider } from "./embeddings.js";
+
+function createProvider(maxInputTokens: number): EmbeddingProvider {
+  return {
+    id: "mock",
+    model: "mock-embed",
+    maxInputTokens,
+    embedQuery: async () => [0],
+    embedBatch: async () => [[0]],
+  };
+}
+
+function createProviderWithoutMaxInputTokens(params: {
+  id: string;
+  model: string;
+}): EmbeddingProvider {
+  return {
+    id: params.id,
+    model: params.model,
+    embedQuery: async () => [0],
+    embedBatch: async () => [[0]],
+  };
+}
+
+type EmbeddingChunks = ReturnType<typeof enforceEmbeddingMaxInputTokens>;
+
+function expectChunksWithinUtf8Bytes(chunks: EmbeddingChunks, maxBytes: number) {
+  const oversized: Array<{ index: number; bytes: number }> = [];
+  for (const [index, chunk] of chunks.entries()) {
+    const bytes = estimateUtf8Bytes(chunk.text);
+    if (bytes > maxBytes) {
+      oversized.push({ index, bytes });
+    }
+  }
+  expect(oversized).toStrictEqual([]);
+}
+
+function expectChunksLineRange(chunks: EmbeddingChunks, startLine: number, endLine: number) {
+  const unexpectedRanges: Array<{ index: number; startLine: number; endLine: number }> = [];
+  for (const [index, chunk] of chunks.entries()) {
+    if (chunk.startLine !== startLine || chunk.endLine !== endLine) {
+      unexpectedRanges.push({ index, startLine: chunk.startLine, endLine: chunk.endLine });
+    }
+  }
+  expect(unexpectedRanges).toStrictEqual([]);
+}
+
+function expectChunksHaveHashes(chunks: EmbeddingChunks) {
+  const invalidHashes: Array<{ index: number; hash: unknown }> = [];
+  for (const [index, chunk] of chunks.entries()) {
+    if (typeof chunk.hash !== "string" || chunk.hash.length === 0) {
+      invalidHashes.push({ index, hash: chunk.hash });
+    }
+  }
+  expect(invalidHashes).toStrictEqual([]);
+}
+
+function joinedChunkText(chunks: EmbeddingChunks): string {
+  let text = "";
+  for (const chunk of chunks) {
+    text += chunk.text;
+  }
+  return text;
+}
+
+describe("embedding chunk limits", () => {
+  it("splits oversized chunks so each embedding input stays <= maxInputTokens bytes", () => {
+    const provider = createProvider(8192);
+    const input = {
+      startLine: 1,
+      endLine: 1,
+      text: "x".repeat(9000),
+      hash: "ignored",
+    };
+
+    const out = enforceEmbeddingMaxInputTokens(provider, [input]);
+    expect(out.length).toBeGreaterThan(1);
+    expect(joinedChunkText(out)).toBe(input.text);
+    expectChunksWithinUtf8Bytes(out, 8192);
+    expectChunksLineRange(out, 1, 1);
+    expectChunksHaveHashes(out);
+  });
+
+  it("does not split inside surrogate pairs (emoji)", () => {
+    const provider = createProvider(8192);
+    const emoji = "😀";
+    const inputText = `${emoji.repeat(2100)}\n${emoji.repeat(2100)}`;
+
+    const out = enforceEmbeddingMaxInputTokens(provider, [
+      { startLine: 1, endLine: 2, text: inputText, hash: "ignored" },
+    ]);
+
+    expect(out.length).toBeGreaterThan(1);
+    expect(joinedChunkText(out)).toBe(inputText);
+    expectChunksWithinUtf8Bytes(out, 8192);
+
+    // If we split inside surrogate pairs we'd likely end up with replacement chars.
+    expect(joinedChunkText(out)).not.toContain("\uFFFD");
+  });
+
+  it("uses conservative fallback limits for local providers without declared maxInputTokens", () => {
+    const provider = createProviderWithoutMaxInputTokens({
+      id: "local",
+      model: "unknown-local-embedding",
+    });
+
+    const out = enforceEmbeddingMaxInputTokens(provider, [
+      {
+        startLine: 1,
+        endLine: 1,
+        text: "x".repeat(3000),
+        hash: "ignored",
+      },
+    ]);
+
+    expect(out.length).toBeGreaterThan(1);
+    expectChunksWithinUtf8Bytes(out, 2048);
+  });
+
+  it("honors hard safety caps lower than provider maxInputTokens", () => {
+    const provider = createProvider(8192);
+    const out = enforceEmbeddingMaxInputTokens(
+      provider,
+      [
+        {
+          startLine: 1,
+          endLine: 1,
+          text: "x".repeat(8100),
+          hash: "ignored",
+        },
+      ],
+      8000,
+    );
+
+    expect(out.length).toBeGreaterThan(1);
+    expectChunksWithinUtf8Bytes(out, 8000);
+  });
+});

packages/memory-host-sdk/src/host/embedding-chunk-limits.ts

@@ -0,0 +1,51 @@
+// Memory Host SDK module implements embedding chunk limits behavior.
+import { estimateUtf8Bytes, splitTextToUtf8ByteLimit } from "./embedding-input-limits.js";
+import { hasNonTextEmbeddingParts } from "./embedding-inputs.js";
+import { resolveEmbeddingMaxInputTokens } from "./embedding-model-limits.js";
+import type { EmbeddingProvider } from "./embeddings.js";
+import { hashText } from "./hash.js";
+import type { MemoryChunk } from "./internal.js";
+
+// Enforces provider byte budgets before chunks reach embedding workers.
+
+/**
+ * Split text-only chunks to the provider's effective input limit.
+ *
+ * Structured multimodal chunks are preserved because only the provider can decide how to count
+ * non-text parts.
+ */
+export function enforceEmbeddingMaxInputTokens(
+  provider: EmbeddingProvider,
+  chunks: MemoryChunk[],
+  hardMaxInputTokens?: number,
+): MemoryChunk[] {
+  const providerMaxInputTokens = resolveEmbeddingMaxInputTokens(provider);
+  const maxInputTokens =
+    typeof hardMaxInputTokens === "number" && hardMaxInputTokens > 0
+      ? Math.min(providerMaxInputTokens, hardMaxInputTokens)
+      : providerMaxInputTokens;
+  const out: MemoryChunk[] = [];
+
+  for (const chunk of chunks) {
+    if (hasNonTextEmbeddingParts(chunk.embeddingInput)) {
+      out.push(chunk);
+      continue;
+    }
+    if (estimateUtf8Bytes(chunk.text) <= maxInputTokens) {
+      out.push(chunk);
+      continue;
+    }
+
+    for (const text of splitTextToUtf8ByteLimit(chunk.text, maxInputTokens)) {
+      out.push({
+        startLine: chunk.startLine,
+        endLine: chunk.endLine,
+        text,
+        hash: hashText(text),
+        embeddingInput: { text },
+      });
+    }
+  }
+
+  return out;
+}

packages/memory-host-sdk/src/host/embedding-defaults.ts

@@ -0,0 +1,5 @@
+// Shared embedding model defaults for builtin memory providers.
+
+/** Default local embedding model used when config omits an explicit model. */
+export const DEFAULT_LOCAL_MODEL =
+  "hf:ggml-org/embeddinggemma-300m-qat-q8_0-GGUF/embeddinggemma-300m-qat-Q8_0.gguf";

packages/memory-host-sdk/src/host/embedding-input-limits.ts

@@ -0,0 +1,86 @@
+// Memory Host SDK module implements embedding input limits behavior.
+import type { EmbeddingInput } from "./embedding-inputs.js";
+
+// Helpers for enforcing embedding model input size limits.
+//
+// We use UTF-8 byte length as a conservative upper bound for tokenizer output.
+// Tokenizers operate over bytes; a token must contain at least one byte, so
+// token_count <= utf8_byte_length.
+
+export function estimateUtf8Bytes(text: string): number {
+  if (!text) {
+    return 0;
+  }
+  return Buffer.byteLength(text, "utf8");
+}
+
+export function estimateStructuredEmbeddingInputBytes(input: EmbeddingInput): number {
+  if (!input.parts?.length) {
+    return estimateUtf8Bytes(input.text);
+  }
+  let total = 0;
+  for (const part of input.parts) {
+    if (part.type === "text") {
+      total += estimateUtf8Bytes(part.text);
+      continue;
+    }
+    total += estimateUtf8Bytes(part.mimeType);
+    total += estimateUtf8Bytes(part.data);
+  }
+  return total;
+}
+
+export function splitTextToUtf8ByteLimit(text: string, maxUtf8Bytes: number): string[] {
+  if (maxUtf8Bytes <= 0) {
+    return [text];
+  }
+  if (estimateUtf8Bytes(text) <= maxUtf8Bytes) {
+    return [text];
+  }
+
+  const parts: string[] = [];
+  let cursor = 0;
+  while (cursor < text.length) {
+    // The number of UTF-16 code units is always <= the number of UTF-8 bytes.
+    // This makes `cursor + maxUtf8Bytes` a safe upper bound on the next split point.
+    let low = cursor + 1;
+    let high = Math.min(text.length, cursor + maxUtf8Bytes);
+    let best = cursor;
+
+    while (low <= high) {
+      const mid = Math.floor((low + high) / 2);
+      const bytes = estimateUtf8Bytes(text.slice(cursor, mid));
+      if (bytes <= maxUtf8Bytes) {
+        best = mid;
+        low = mid + 1;
+      } else {
+        high = mid - 1;
+      }
+    }
+
+    if (best <= cursor) {
+      best = Math.min(text.length, cursor + 1);
+    }
+
+    // Avoid splitting inside a surrogate pair.
+    if (
+      best < text.length &&
+      best > cursor &&
+      text.charCodeAt(best - 1) >= 0xd800 &&
+      text.charCodeAt(best - 1) <= 0xdbff &&
+      text.charCodeAt(best) >= 0xdc00 &&
+      text.charCodeAt(best) <= 0xdfff
+    ) {
+      best -= 1;
+    }
+
+    const part = text.slice(cursor, best);
+    if (!part) {
+      break;
+    }
+    parts.push(part);
+    cursor = best;
+  }
+
+  return parts;
+}

packages/memory-host-sdk/src/host/embedding-inputs.ts

@@ -0,0 +1,43 @@
+// Public embedding input contract for text and inline multimodal parts.
+
+/** Text part passed through embedding providers that support structured input. */
+export type EmbeddingInputTextPart = {
+  type: "text";
+  text: string;
+};
+
+/** Inline binary payload encoded for providers with multimodal embedding support. */
+export type EmbeddingInputInlineDataPart = {
+  type: "inline-data";
+  mimeType: string;
+  data: string;
+};
+
+/** Single structured embedding input part. */
+export type EmbeddingInputPart = EmbeddingInputTextPart | EmbeddingInputInlineDataPart;
+
+/** Provider-facing input while preserving the plain text fallback. */
+export type EmbeddingInput = {
+  text: string;
+  parts?: EmbeddingInputPart[];
+};
+
+/** Build the common text-only embedding input shape. */
+export function buildTextEmbeddingInput(text: string): EmbeddingInput {
+  return { text };
+}
+
+/** Narrow an embedding part to an inline-data payload. */
+export function isInlineDataEmbeddingInputPart(
+  part: EmbeddingInputPart,
+): part is EmbeddingInputInlineDataPart {
+  return part.type === "inline-data";
+}
+
+/** Return true when a chunk needs structured provider handling, not text splitting. */
+export function hasNonTextEmbeddingParts(input: EmbeddingInput | undefined): boolean {
+  if (!input?.parts?.length) {
+    return false;
+  }
+  return input.parts.some((part) => isInlineDataEmbeddingInputPart(part));
+}

packages/memory-host-sdk/src/host/embedding-model-limits.ts

@@ -0,0 +1,20 @@
+// Memory Host SDK module implements embedding model limits behavior.
+import type { EmbeddingProvider } from "./embeddings.js";
+
+// Provider input limits are byte-based approximations for pre-embedding chunk splitting.
+
+const DEFAULT_EMBEDDING_MAX_INPUT_TOKENS = 8192;
+const DEFAULT_LOCAL_EMBEDDING_MAX_INPUT_TOKENS = 2048;
+
+/** Resolve the effective embedding input limit for a provider. */
+export function resolveEmbeddingMaxInputTokens(provider: EmbeddingProvider): number {
+  if (typeof provider.maxInputTokens === "number") {
+    return provider.maxInputTokens;
+  }
+
+  if (provider.id === "local") {
+    return DEFAULT_LOCAL_EMBEDDING_MAX_INPUT_TOKENS;
+  }
+
+  return DEFAULT_EMBEDDING_MAX_INPUT_TOKENS;
+}

packages/memory-host-sdk/src/host/embedding-provider-adapter-utils.ts

@@ -0,0 +1,35 @@
+// Memory Host SDK helper module supports embedding provider adapter utils behavior.
+import { normalizeLowercaseStringOrEmpty } from "./string-utils.js";
+
+// Adapter helpers shared by remote embedding provider implementations.
+
+/** Detect missing API key errors from provider auth resolution. */
+export function isMissingEmbeddingApiKeyError(err: unknown): boolean {
+  return err instanceof Error && err.message.includes("No API key found for provider");
+}
+
+/** Return stable cache headers after removing provider-specific secret headers. */
+export function sanitizeEmbeddingCacheHeaders(
+  headers: Record<string, string>,
+  excludedHeaderNames: string[],
+): Array<[string, string]> {
+  const excluded = new Set(
+    excludedHeaderNames.map((name) => normalizeLowercaseStringOrEmpty(name)),
+  );
+  return Object.entries(headers)
+    .filter(([key]) => !excluded.has(normalizeLowercaseStringOrEmpty(key)))
+    .toSorted(([a], [b]) => a.localeCompare(b))
+    .map(([key, value]) => [key, value]);
+}
+
+/** Convert custom-id keyed batch embeddings back to request-index order. */
+export function mapBatchEmbeddingsByIndex(
+  byCustomId: Map<string, number[]>,
+  count: number,
+): number[][] {
+  const embeddings: number[][] = [];
+  for (let index = 0; index < count; index += 1) {
+    embeddings.push(byCustomId.get(String(index)) ?? []);
+  }
+  return embeddings;
+}

packages/memory-host-sdk/src/host/embedding-vectors.ts

@@ -0,0 +1,11 @@
+// Vector normalization helpers used before embedding similarity search.
+
+/** Replace invalid coordinates and L2-normalize non-empty vectors. */
+export function sanitizeAndNormalizeEmbedding(vec: number[]): number[] {
+  const sanitized = vec.map((value) => (Number.isFinite(value) ? value : 0));
+  const magnitude = Math.sqrt(sanitized.reduce((sum, value) => sum + value * value, 0));
+  if (magnitude < 1e-10) {
+    return sanitized;
+  }
+  return sanitized.map((value) => value / magnitude);
+}

packages/memory-host-sdk/src/host/embedding-worker-errors.ts

@@ -0,0 +1,41 @@
+// Typed local embedding worker failures for process and IPC lifecycle handling.
+
+/** Stable error codes emitted by the local embedding worker supervisor. */
+export const LOCAL_EMBEDDING_WORKER_ERROR_CODES = {
+  exited: "LOCAL_EMBEDDING_WORKER_EXITED",
+  processError: "LOCAL_EMBEDDING_WORKER_PROCESS_ERROR",
+  ipcError: "LOCAL_EMBEDDING_WORKER_IPC_ERROR",
+} as const;
+
+/** Error code union for local embedding worker failures. */
+export type LocalEmbeddingWorkerFailureCode =
+  (typeof LOCAL_EMBEDDING_WORKER_ERROR_CODES)[keyof typeof LOCAL_EMBEDDING_WORKER_ERROR_CODES];
+
+/** Cause category for local embedding worker failures. */
+export type LocalEmbeddingWorkerFailureReason = "exit" | "signal" | "process-error" | "ipc";
+
+/** Error shape used by callers that need retry/status decisions. */
+export type LocalEmbeddingWorkerFailureError = Error & {
+  code: LocalEmbeddingWorkerFailureCode;
+  reason: LocalEmbeddingWorkerFailureReason;
+  exitCode?: number | null;
+  signal?: NodeJS.Signals | null;
+};
+
+/** Create a local embedding worker failure with stable metadata fields. */
+export function createLocalEmbeddingWorkerFailureError(params: {
+  message: string;
+  code: LocalEmbeddingWorkerFailureCode;
+  reason: LocalEmbeddingWorkerFailureReason;
+  exitCode?: number | null;
+  signal?: NodeJS.Signals | null;
+  cause?: unknown;
+}): LocalEmbeddingWorkerFailureError {
+  return Object.assign(new Error(params.message), {
+    code: params.code,
+    reason: params.reason,
+    ...(params.exitCode !== undefined ? { exitCode: params.exitCode } : {}),
+    ...(params.signal !== undefined ? { signal: params.signal } : {}),
+    ...(params.cause !== undefined ? { cause: params.cause } : {}),
+  });
+}

packages/memory-host-sdk/src/host/embeddings-debug.ts

@@ -0,0 +1,28 @@
+// Memory Host SDK module implements embeddings debug behavior.
+import { normalizeLowercaseStringOrEmpty } from "./string-utils.js";
+
+// Lightweight stderr debug logging for memory embedding internals.
+
+const debugEmbeddings = isTruthyEnvValue(process.env.OPENCLAW_DEBUG_MEMORY_EMBEDDINGS);
+
+/** Write embedding debug metadata when OPENCLAW_DEBUG_MEMORY_EMBEDDINGS is enabled. */
+export function debugEmbeddingsLog(message: string, meta?: Record<string, unknown>): void {
+  if (!debugEmbeddings) {
+    return;
+  }
+  const suffix = meta ? ` ${JSON.stringify(meta)}` : "";
+  process.stderr.write(`${message}${suffix}\n`);
+}
+
+/** Parse common truthy env values for debug toggles. */
+function isTruthyEnvValue(value?: string): boolean {
+  switch (normalizeLowercaseStringOrEmpty(value)) {
+    case "1":
+    case "on":
+    case "true":
+    case "yes":
+      return true;
+    default:
+      return false;
+  }
+}

packages/memory-host-sdk/src/host/embeddings-model-normalize.test.ts

@@ -0,0 +1,35 @@
+// Memory Host SDK tests cover embeddings model normalize behavior.
+import { describe, expect, it } from "vitest";
+import { normalizeEmbeddingModelWithPrefixes } from "./embeddings-model-normalize.js";
+
+describe("normalizeEmbeddingModelWithPrefixes", () => {
+  it("returns default model when input is blank", () => {
+    expect(
+      normalizeEmbeddingModelWithPrefixes({
+        model: "   ",
+        defaultModel: "fallback-model",
+        prefixes: ["openai/"],
+      }),
+    ).toBe("fallback-model");
+  });
+
+  it("strips the first matching prefix", () => {
+    expect(
+      normalizeEmbeddingModelWithPrefixes({
+        model: "openai/text-embedding-3-small",
+        defaultModel: "fallback-model",
+        prefixes: ["openai/"],
+      }),
+    ).toBe("text-embedding-3-small");
+  });
+
+  it("keeps explicit model names when no prefix matches", () => {
+    expect(
+      normalizeEmbeddingModelWithPrefixes({
+        model: "voyage-4-large",
+        defaultModel: "fallback-model",
+        prefixes: ["voyage/"],
+      }),
+    ).toBe("voyage-4-large");
+  });
+});

packages/memory-host-sdk/src/host/embeddings-model-normalize.ts

@@ -0,0 +1,19 @@
+// Normalizes user-provided embedding model ids by removing accepted provider prefixes.
+
+/** Trim a configured model id, fall back when empty, and strip known prefixes. */
+export function normalizeEmbeddingModelWithPrefixes(params: {
+  model: string;
+  defaultModel: string;
+  prefixes: string[];
+}): string {
+  const trimmed = params.model.trim();
+  if (!trimmed) {
+    return params.defaultModel;
+  }
+  for (const prefix of params.prefixes) {
+    if (trimmed.startsWith(prefix)) {
+      return trimmed.slice(prefix.length);
+    }
+  }
+  return trimmed;
+}

packages/memory-host-sdk/src/host/embeddings-remote-client.test.ts

@@ -0,0 +1,57 @@
+// Memory Host SDK tests cover embeddings remote client behavior.
+import { describe, expect, it, vi } from "vitest";
+import { resolveRemoteEmbeddingBearerClient } from "./embeddings-remote-client.js";
+
+describe("resolveRemoteEmbeddingBearerClient", () => {
+  it("uses configured OpenAI provider baseUrl for memory embeddings", async () => {
+    const client = await resolveRemoteEmbeddingBearerClient({
+      provider: "openai",
+      defaultBaseUrl: "https://api.openai.com/v1",
+      options: {
+        agentDir: "/tmp/openclaw-agent",
+        config: {
+          models: {
+            providers: {
+              openai: {
+                baseUrl: "https://proxy.example.test/openai/v1",
+              },
+            },
+          },
+        } as never,
+        model: "text-embedding-3-small",
+        remote: {
+          apiKey: "sk-test",
+        },
+      },
+    });
+
+    expect(client.baseUrl).toBe("https://proxy.example.test/openai/v1");
+  });
+
+  it("adds OpenClaw attribution to native OpenAI embedding requests", async () => {
+    vi.stubEnv("OPENCLAW_VERSION", "2026.3.22");
+    const client = await resolveRemoteEmbeddingBearerClient({
+      provider: "openai",
+      defaultBaseUrl: "https://api.openai.com/v1",
+      options: {
+        config: { models: {} } as never,
+        model: "text-embedding-3-large",
+        remote: {
+          apiKey: "sk-test",
+          headers: {
+            originator: "openclaw",
+            "User-Agent": "openclaw",
+          },
+        },
+      },
+    });
+
+    expect(client.headers).toEqual({
+      Authorization: "Bearer sk-test",
+      "Content-Type": "application/json",
+      originator: "openclaw",
+      version: "2026.3.22",
+      "User-Agent": "openclaw/2026.3.22",
+    });
+  });
+});

packages/memory-host-sdk/src/host/embeddings-remote-client.ts

@@ -0,0 +1,71 @@
+// Memory Host SDK module implements embeddings remote client behavior.
+import type { EmbeddingProviderOptions } from "./embeddings.types.js";
+import { requireApiKey, resolveApiKeyForProvider } from "./openclaw-runtime-auth.js";
+import { buildRemoteBaseUrlPolicy } from "./remote-http.js";
+import { resolveMemorySecretInputString } from "./secret-input.js";
+import type { SsrFPolicy } from "./ssrf-policy.js";
+import { normalizeOptionalString } from "./string-utils.js";
+
+// Builds authenticated remote embedding HTTP clients from agent memory config.
+
+/** Provider id used for remote embedding auth and config lookup. */
+export type RemoteEmbeddingProviderId = string;
+
+/** Attribution headers for native OpenAI embedding calls. */
+function resolveOpenClawAttributionHeaders(): Record<string, string> {
+  const version = typeof process !== "undefined" ? process.env.OPENCLAW_VERSION?.trim() : undefined;
+  return {
+    originator: "openclaw",
+    ...(version ? { version } : {}),
+    "User-Agent": version ? `openclaw/${version}` : "openclaw",
+  };
+}
+
+/** Detect the native OpenAI embeddings API route that accepts attribution headers. */
+function isNativeOpenAIEmbeddingRoute(provider: string, baseUrl: string): boolean {
+  if (provider !== "openai") {
+    return false;
+  }
+  try {
+    return new URL(baseUrl).hostname.toLowerCase().replace(/\.+$/, "") === "api.openai.com";
+  } catch {
+    return false;
+  }
+}
+
+/** Resolve base URL, bearer headers, header overrides, and SSRF policy for remote embeddings. */
+export async function resolveRemoteEmbeddingBearerClient(params: {
+  provider: RemoteEmbeddingProviderId;
+  options: EmbeddingProviderOptions;
+  defaultBaseUrl: string;
+}): Promise<{ baseUrl: string; headers: Record<string, string>; ssrfPolicy?: SsrFPolicy }> {
+  const remote = params.options.remote;
+  const remoteApiKey = resolveMemorySecretInputString({
+    value: remote?.apiKey,
+    path: "agents.*.memorySearch.remote.apiKey",
+  });
+  const remoteBaseUrl = normalizeOptionalString(remote?.baseUrl);
+  const providerConfig = params.options.config.models?.providers?.[params.provider];
+  const apiKey = remoteApiKey
+    ? remoteApiKey
+    : requireApiKey(
+        await resolveApiKeyForProvider({
+          provider: params.provider,
+          cfg: params.options.config,
+          agentDir: params.options.agentDir,
+        }),
+        params.provider,
+      );
+  const baseUrl =
+    remoteBaseUrl || normalizeOptionalString(providerConfig?.baseUrl) || params.defaultBaseUrl;
+  const headerOverrides = Object.assign({}, providerConfig?.headers, remote?.headers);
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    Authorization: `Bearer ${apiKey}`,
+    ...headerOverrides,
+  };
+  if (isNativeOpenAIEmbeddingRoute(params.provider, baseUrl)) {
+    Object.assign(headers, resolveOpenClawAttributionHeaders());
+  }
+  return { baseUrl, headers, ssrfPolicy: buildRemoteBaseUrlPolicy(baseUrl) };
+}

packages/memory-host-sdk/src/host/embeddings-remote-fetch.test.ts

@@ -0,0 +1,141 @@
+// Memory Host SDK tests cover embeddings remote fetch behavior.
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import { fetchRemoteEmbeddingVectors } from "./embeddings-remote-fetch.js";
+
+const postJsonMock = vi.hoisted(() => vi.fn());
+
+vi.mock("./post-json.js", () => ({
+  postJson: postJsonMock,
+}));
+
+function requirePostJsonParams(): {
+  url?: unknown;
+  headers?: unknown;
+  signal?: unknown;
+  body?: unknown;
+  errorPrefix?: unknown;
+} {
+  const [call] = postJsonMock.mock.calls;
+  if (!call) {
+    throw new Error("expected postJson call");
+  }
+  const [params] = call;
+  if (typeof params !== "object" || params === null || Array.isArray(params)) {
+    throw new Error("expected postJson params to be an object");
+  }
+  return params;
+}
+
+describe("fetchRemoteEmbeddingVectors", () => {
+  beforeEach(() => {
+    postJsonMock.mockReset();
+  });
+
+  it("maps remote embedding response data to vectors", async () => {
+    postJsonMock.mockImplementationOnce(async (params) => {
+      return await params.parse({
+        data: [{ embedding: [0.1, 0.2] }, { embedding: [0.4] }, { embedding: [0.3] }],
+      });
+    });
+
+    const vectors = await fetchRemoteEmbeddingVectors({
+      url: "https://memory.example/v1/embeddings",
+      headers: { Authorization: "Bearer test" },
+      body: { input: ["one", "two", "three"] },
+      errorPrefix: "embedding fetch failed",
+    });
+
+    expect(vectors).toEqual([[0.1, 0.2], [0.4], [0.3]]);
+    const postJsonParams = requirePostJsonParams();
+    expect(postJsonParams.url).toBe("https://memory.example/v1/embeddings");
+    expect(postJsonParams.headers).toEqual({ Authorization: "Bearer test" });
+    expect(postJsonParams.body).toEqual({ input: ["one", "two", "three"] });
+    expect(postJsonParams.errorPrefix).toBe("embedding fetch failed");
+  });
+
+  it("passes abort signals to the JSON request", async () => {
+    const controller = new AbortController();
+    postJsonMock.mockImplementationOnce(async (params) => {
+      return await params.parse({ data: [{ embedding: [0.1] }] });
+    });
+
+    await fetchRemoteEmbeddingVectors({
+      url: "https://memory.example/v1/embeddings",
+      headers: {},
+      signal: controller.signal,
+      body: { input: ["one"] },
+      errorPrefix: "embedding fetch failed",
+    });
+
+    expect(requirePostJsonParams().signal).toBe(controller.signal);
+  });
+
+  it("throws a status-rich error on non-ok responses", async () => {
+    postJsonMock.mockRejectedValueOnce(new Error("embedding fetch failed: 403 forbidden"));
+
+    await expect(
+      fetchRemoteEmbeddingVectors({
+        url: "https://memory.example/v1/embeddings",
+        headers: {},
+        body: { input: ["one"] },
+        errorPrefix: "embedding fetch failed",
+      }),
+    ).rejects.toThrow("embedding fetch failed: 403 forbidden");
+  });
+
+  it("rejects non-object embedding responses", async () => {
+    postJsonMock.mockImplementationOnce(async (params) => await params.parse([]));
+
+    await expect(
+      fetchRemoteEmbeddingVectors({
+        url: "https://memory.example/v1/embeddings",
+        headers: {},
+        body: { input: ["one"] },
+        errorPrefix: "embedding fetch failed",
+      }),
+    ).rejects.toThrow("embedding fetch failed: malformed JSON response");
+  });
+
+  it("rejects missing embedding data arrays", async () => {
+    postJsonMock.mockImplementationOnce(async (params) => await params.parse({}));
+
+    await expect(
+      fetchRemoteEmbeddingVectors({
+        url: "https://memory.example/v1/embeddings",
+        headers: {},
+        body: { input: ["one"] },
+        errorPrefix: "embedding fetch failed",
+      }),
+    ).rejects.toThrow("embedding fetch failed: malformed JSON response");
+  });
+
+  it("rejects embedding counts that do not match the submitted input batch", async () => {
+    postJsonMock.mockImplementationOnce(async (params) => {
+      return await params.parse({ data: [{ embedding: [0.1] }] });
+    });
+
+    await expect(
+      fetchRemoteEmbeddingVectors({
+        url: "https://memory.example/v1/embeddings",
+        headers: {},
+        body: { input: ["one", "two"] },
+        errorPrefix: "embedding fetch failed",
+      }),
+    ).rejects.toThrow("embedding fetch failed: malformed JSON response");
+  });
+
+  it("rejects wrong nested embedding vector types", async () => {
+    postJsonMock.mockImplementationOnce(async (params) => {
+      return await params.parse({ data: [{ embedding: [0.1, "bad"] }] });
+    });
+
+    await expect(
+      fetchRemoteEmbeddingVectors({
+        url: "https://memory.example/v1/embeddings",
+        headers: {},
+        body: { input: ["one"] },
+        errorPrefix: "embedding fetch failed",
+      }),
+    ).rejects.toThrow("embedding fetch failed: malformed JSON response");
+  });
+});

packages/memory-host-sdk/src/host/embeddings-remote-fetch.ts

@@ -0,0 +1,74 @@
+// Memory Host SDK module implements embeddings remote fetch behavior.
+import { postJson } from "./post-json.js";
+import type { SsrFPolicy } from "./ssrf-policy.js";
+
+// Fetches and validates OpenAI-compatible embedding responses.
+
+/** Narrow unknown JSON payloads to plain objects. */
+function asRecord(value: unknown): Record<string, unknown> | undefined {
+  return typeof value === "object" && value !== null && !Array.isArray(value)
+    ? (value as Record<string, unknown>)
+    : undefined;
+}
+
+/** Build the common malformed embedding response error. */
+function malformedEmbeddingResponse(errorPrefix: string): Error {
+  return new Error(`${errorPrefix}: malformed JSON response`);
+}
+
+/** Validate and return one finite embedding vector. */
+function readEmbeddingVector(value: unknown, errorPrefix: string): number[] {
+  if (!Array.isArray(value)) {
+    throw malformedEmbeddingResponse(errorPrefix);
+  }
+  for (const entry of value) {
+    if (typeof entry !== "number" || !Number.isFinite(entry)) {
+      throw malformedEmbeddingResponse(errorPrefix);
+    }
+  }
+  return value;
+}
+
+/** Resolve expected response count from the request body when input is an array. */
+function resolveExpectedEmbeddingCount(body: unknown): number | undefined {
+  const input = asRecord(body)?.input;
+  return Array.isArray(input) ? input.length : undefined;
+}
+
+/** POST an embedding request and return validated vectors in provider response order. */
+export async function fetchRemoteEmbeddingVectors(params: {
+  url: string;
+  headers: Record<string, string>;
+  ssrfPolicy?: SsrFPolicy;
+  fetchImpl?: typeof fetch;
+  signal?: AbortSignal;
+  body: unknown;
+  errorPrefix: string;
+}): Promise<number[][]> {
+  return await postJson({
+    url: params.url,
+    headers: params.headers,
+    ssrfPolicy: params.ssrfPolicy,
+    fetchImpl: params.fetchImpl,
+    signal: params.signal,
+    body: params.body,
+    errorPrefix: params.errorPrefix,
+    parse: (payload) => {
+      const root = asRecord(payload);
+      if (!root || !Array.isArray(root.data)) {
+        throw malformedEmbeddingResponse(params.errorPrefix);
+      }
+      const expectedCount = resolveExpectedEmbeddingCount(params.body);
+      if (expectedCount !== undefined && root.data.length !== expectedCount) {
+        throw malformedEmbeddingResponse(params.errorPrefix);
+      }
+      return root.data.map((entry) => {
+        const record = asRecord(entry);
+        if (!record) {
+          throw malformedEmbeddingResponse(params.errorPrefix);
+        }
+        return readEmbeddingVector(record.embedding, params.errorPrefix);
+      });
+    },
+  });
+}

packages/memory-host-sdk/src/host/embeddings-remote-provider.ts

@@ -0,0 +1,72 @@
+// Memory Host SDK module implements embeddings remote provider behavior.
+import {
+  resolveRemoteEmbeddingBearerClient,
+  type RemoteEmbeddingProviderId,
+} from "./embeddings-remote-client.js";
+import { fetchRemoteEmbeddingVectors } from "./embeddings-remote-fetch.js";
+import type { EmbeddingProvider, EmbeddingProviderOptions } from "./embeddings.types.js";
+import type { SsrFPolicy } from "./ssrf-policy.js";
+
+// Remote embedding provider factory for OpenAI-compatible embeddings APIs.
+
+/** HTTP client details required by a remote embedding provider. */
+export type RemoteEmbeddingClient = {
+  baseUrl: string;
+  headers: Record<string, string>;
+  ssrfPolicy?: SsrFPolicy;
+  fetchImpl?: typeof fetch;
+  model: string;
+};
+
+/** Create an EmbeddingProvider backed by a remote embeddings endpoint. */
+export function createRemoteEmbeddingProvider(params: {
+  id: string;
+  client: RemoteEmbeddingClient;
+  errorPrefix: string;
+  maxInputTokens?: number;
+}): EmbeddingProvider {
+  const { client } = params;
+  const url = `${client.baseUrl.replace(/\/$/, "")}/embeddings`;
+
+  const embed = async (input: string[], signal?: AbortSignal): Promise<number[][]> => {
+    if (input.length === 0) {
+      return [];
+    }
+    return await fetchRemoteEmbeddingVectors({
+      url,
+      headers: client.headers,
+      ssrfPolicy: client.ssrfPolicy,
+      fetchImpl: client.fetchImpl,
+      signal,
+      body: { model: client.model, input },
+      errorPrefix: params.errorPrefix,
+    });
+  };
+
+  return {
+    id: params.id,
+    model: client.model,
+    ...(typeof params.maxInputTokens === "number" ? { maxInputTokens: params.maxInputTokens } : {}),
+    embedQuery: async (text, options) => {
+      const [vec] = await embed([text], options?.signal);
+      return vec ?? [];
+    },
+    embedBatch: async (texts, options) => await embed(texts, options?.signal),
+  };
+}
+
+/** Resolve a normalized remote embedding client from provider config and model options. */
+export async function resolveRemoteEmbeddingClient(params: {
+  provider: RemoteEmbeddingProviderId;
+  options: EmbeddingProviderOptions;
+  defaultBaseUrl: string;
+  normalizeModel: (model: string) => string;
+}): Promise<RemoteEmbeddingClient> {
+  const { baseUrl, headers, ssrfPolicy } = await resolveRemoteEmbeddingBearerClient({
+    provider: params.provider,
+    options: params.options,
+    defaultBaseUrl: params.defaultBaseUrl,
+  });
+  const model = params.normalizeModel(params.options.model);
+  return { baseUrl, headers, ssrfPolicy, model };
+}

packages/memory-host-sdk/src/host/embeddings-worker-child.ts

@@ -0,0 +1,120 @@
+// Memory Host SDK module implements embeddings worker child behavior.
+import { createLocalEmbeddingProviderInProcess } from "./embeddings.js";
+import type { EmbeddingProvider, EmbeddingProviderOptions } from "./embeddings.types.js";
+
+// Child process entrypoint for local embedding work.
+
+/** Request payloads accepted from the parent worker client. */
+type LocalEmbeddingWorkerRequest =
+  | {
+      id: number;
+      type: "initialize";
+      options: EmbeddingProviderOptions;
+    }
+  | {
+      id: number;
+      type: "embedQuery";
+      options: EmbeddingProviderOptions;
+      text: string;
+    }
+  | {
+      id: number;
+      type: "embedBatch";
+      options: EmbeddingProviderOptions;
+      texts: string[];
+    }
+  | {
+      id: number;
+      type: "close";
+    };
+
+/** Serialized error shape returned over JSON IPC. */
+type LocalEmbeddingWorkerSerializedError = {
+  message: string;
+  code?: string;
+};
+
+let provider: EmbeddingProvider | null = null;
+let providerOptionsKey: string | null = null;
+let requestQueue: Promise<void> = Promise.resolve();
+
+/** Send one JSON IPC message when the child still has an IPC channel. */
+function send(message: unknown): void {
+  if (typeof process.send === "function") {
+    process.send(message);
+  }
+}
+
+/** Reuse the current provider while options are unchanged, otherwise rebuild it. */
+async function getProvider(options: EmbeddingProviderOptions): Promise<EmbeddingProvider> {
+  const key = JSON.stringify(options);
+  if (provider && providerOptionsKey === key) {
+    return provider;
+  }
+  await provider?.close?.();
+  provider = await createLocalEmbeddingProviderInProcess(options);
+  providerOptionsKey = key;
+  return provider;
+}
+
+/** Close and forget the active in-process provider. */
+async function closeProvider(): Promise<void> {
+  const current = provider;
+  provider = null;
+  providerOptionsKey = null;
+  await current?.close?.();
+}
+
+/** Preserve error message and code across JSON IPC. */
+function serializeError(err: unknown): LocalEmbeddingWorkerSerializedError {
+  if (!(err instanceof Error)) {
+    return { message: String(err) };
+  }
+  const code = (err as Error & { code?: unknown }).code;
+  return {
+    message: err.message,
+    ...(typeof code === "string" ? { code } : {}),
+  };
+}
+
+/** Handle one parent request after queue serialization. */
+async function handleRequest(request: LocalEmbeddingWorkerRequest): Promise<void> {
+  if (request.type === "close") {
+    await closeProvider();
+    send({ id: request.id, ok: true });
+    return;
+  }
+
+  const currentProvider = await getProvider(request.options);
+  if (request.type === "initialize") {
+    send({ id: request.id, ok: true });
+    return;
+  }
+  if (request.type === "embedQuery") {
+    const value = await currentProvider.embedQuery(request.text);
+    send({ id: request.id, ok: true, value });
+    return;
+  }
+
+  const value = await currentProvider.embedBatch(request.texts);
+  send({ id: request.id, ok: true, value });
+}
+
+// Requests are serialized so node-llama-cpp context state is not used concurrently.
+process.on("message", (message) => {
+  const request = message as LocalEmbeddingWorkerRequest;
+  requestQueue = requestQueue.then(async () => {
+    try {
+      await handleRequest(request);
+    } catch (err) {
+      send({ id: request.id, ok: false, error: serializeError(err) });
+    }
+  });
+});
+
+// Parent disconnect means the worker is orphaned; close provider resources before exiting.
+process.once("disconnect", () => {
+  void closeProvider().finally(() => {
+    process.exit(0);
+  });
+});

packages/memory-host-sdk/src/host/embeddings-worker.ts

@@ -0,0 +1,413 @@
+// Memory Host SDK module implements embeddings worker behavior.
+import { fork, type ChildProcess } from "node:child_process";
+import path from "node:path";
+import { fileURLToPath } from "node:url";
+import { DEFAULT_LOCAL_MODEL } from "./embedding-defaults.js";
+import {
+  createLocalEmbeddingWorkerFailureError,
+  LOCAL_EMBEDDING_WORKER_ERROR_CODES,
+} from "./embedding-worker-errors.js";
+import type { LocalEmbeddingProviderRuntimeOptions } from "./embeddings.js";
+import type {
+  EmbeddingProvider,
+  EmbeddingProviderCallOptions,
+  EmbeddingProviderOptions,
+} from "./embeddings.types.js";
+import { normalizeOptionalString } from "./string-utils.js";
+
+// Parent-side local embedding worker client for isolating node-llama-cpp state.
+
+/** Request payloads sent from the parent process to the local embedding worker child. */
+type LocalEmbeddingWorkerRequestPayload =
+  | {
+      type: "initialize";
+      options: EmbeddingProviderOptions;
+    }
+  | {
+      type: "embedQuery";
+      options: EmbeddingProviderOptions;
+      text: string;
+    }
+  | {
+      type: "embedBatch";
+      options: EmbeddingProviderOptions;
+      texts: string[];
+    }
+  | {
+      type: "close";
+    };
+
+type LocalEmbeddingWorkerRequest = LocalEmbeddingWorkerRequestPayload & { id: number };
+
+/** Response payloads sent from the local embedding worker child back to the parent. */
+type LocalEmbeddingWorkerResponse =
+  | {
+      id: number;
+      ok: true;
+      value?: number[] | number[][];
+    }
+  | {
+      id: number;
+      ok: false;
+      error:
+        | string
+        | {
+            message?: string;
+            code?: string;
+          };
+    };
+
+/** Pending parent request plus abort cleanup. */
+type PendingRequest = {
+  resolve: (value: number[] | number[][] | undefined) => void;
+  reject: (err: unknown) => void;
+  abort?: () => void;
+};
+
+/** Resolve the worker child script for source, package, and bundled runtime layouts. */
+function resolveDefaultWorkerScriptPath(): string {
+  const currentPath = fileURLToPath(import.meta.url);
+  const extension = path.extname(currentPath);
+  const currentName = path.basename(currentPath);
+  const sibling =
+    extension === ".ts"
+      ? "embeddings-worker-child.ts"
+      : currentName.startsWith("embeddings-worker.")
+        ? "embeddings-worker-child.js"
+        : "memory-core-local-embedding-worker.js";
+  return path.join(path.dirname(currentPath), sibling);
+}
+
+/** Keep only local embedding options that are safe and necessary to send over IPC. */
+function serializeLocalEmbeddingOptions(
+  options: EmbeddingProviderOptions,
+  runtimeOptions?: LocalEmbeddingProviderRuntimeOptions,
+): EmbeddingProviderOptions {
+  return {
+    config: {},
+    provider: "local",
+    model: options.model,
+    fallback: "none",
+    outputDimensionality: options.outputDimensionality,
+    local: {
+      ...options.local,
+      ...(runtimeOptions?.nodeLlamaCppImportUrl
+        ? { nodeLlamaCppImportUrl: runtimeOptions.nodeLlamaCppImportUrl }
+        : {}),
+    } as EmbeddingProviderOptions["local"],
+  };
+}
+
+/** Create a typed failure for unexpected worker process exits. */
+function createWorkerExitError(code: number | null, signal: NodeJS.Signals | null): Error {
+  const detail = signal ? `signal ${signal}` : `exit code ${code ?? "unknown"}`;
+  return createLocalEmbeddingWorkerFailureError({
+    message: `Local embedding worker exited unexpectedly (${detail})`,
+    code: LOCAL_EMBEDDING_WORKER_ERROR_CODES.exited,
+    reason: signal ? "signal" : "exit",
+    exitCode: code,
+    signal,
+  });
+}
+
+/** Convert worker response errors into Error objects while preserving worker error codes. */
+function createWorkerResponseError(error: LocalEmbeddingWorkerResponse & { ok: false }): Error {
+  if (typeof error.error === "object" && error.error) {
+    const message = error.error.message || "Local embedding worker failed";
+    const workerError = new Error(message) as Error & { code?: string };
+    if (error.error.code) {
+      workerError.code = error.error.code;
+    }
+    return workerError;
+  }
+  return new Error(error.error || "Local embedding worker failed");
+}
+
+const WORKER_UNSAFE_EXEC_ARGV_FLAGS = new Set(["--inspect", "--inspect-brk"]);
+
+const WORKER_UNSAFE_EXEC_ARGV_FLAGS_WITH_VALUE = new Set([
+  "--eval",
+  "-e",
+  "--print",
+  "-p",
+  "--input-type",
+  "--inspect-port",
+]);
+
+const WORKER_UNSAFE_EXEC_ARGV_OPTION_PREFIXES = [
+  "--eval=",
+  "--print=",
+  "--input-type=",
+  "--inspect=",
+  "--inspect-brk=",
+  "--inspect-port=",
+];
+
+const WORKER_CLOSE_GRACE_MS = 250;
+
+/** Drop execArgv flags that would make forked workers debug/eval stateful or unsafe. */
+function resolveWorkerExecArgv(): string[] {
+  const args: string[] = [];
+  let skipNext = false;
+  for (const arg of process.execArgv) {
+    if (skipNext) {
+      skipNext = false;
+      continue;
+    }
+    if (WORKER_UNSAFE_EXEC_ARGV_FLAGS.has(arg)) {
+      continue;
+    }
+    if (WORKER_UNSAFE_EXEC_ARGV_FLAGS_WITH_VALUE.has(arg)) {
+      skipNext = true;
+      continue;
+    }
+    if (WORKER_UNSAFE_EXEC_ARGV_OPTION_PREFIXES.some((prefix) => arg.startsWith(prefix))) {
+      continue;
+    }
+    args.push(arg);
+  }
+  return args;
+}
+
+/** IPC client that serializes local embedding calls through one child process. */
+class LocalEmbeddingWorkerClient {
+  private child: ChildProcess | null = null;
+  private nextRequestId = 1;
+  private pending = new Map<number, PendingRequest>();
+
+  constructor(private readonly scriptPath: string) {}
+
+  /** Start or reuse the child worker and initialize its provider. */
+  async initialize(options: EmbeddingProviderOptions): Promise<void> {
+    await this.send({ type: "initialize", options });
+  }
+
+  /** Request one query embedding from the child worker. */
+  async embedQuery(
+    options: EmbeddingProviderOptions,
+    text: string,
+    callOptions?: EmbeddingProviderCallOptions,
+  ): Promise<number[]> {
+    const result = await this.send({ type: "embedQuery", options, text }, callOptions);
+    return Array.isArray(result) ? (result as number[]) : [];
+  }
+
+  /** Request a batch of embeddings from the child worker. */
+  async embedBatch(
+    options: EmbeddingProviderOptions,
+    texts: string[],
+    callOptions?: EmbeddingProviderCallOptions,
+  ): Promise<number[][]> {
+    const result = await this.send({ type: "embedBatch", options, texts }, callOptions);
+    return Array.isArray(result) ? (result as number[][]) : [];
+  }
+
+  /** Ask the child to close gracefully, then force shutdown after a short grace period. */
+  async close(): Promise<void> {
+    const child = this.child;
+    if (!child) {
+      return;
+    }
+    let timeout: NodeJS.Timeout | undefined;
+    const closeRequest = this.send({ type: "close" }).then(() => "closed" as const);
+    const closeTimeout = new Promise<"timeout">((resolve) => {
+      timeout = setTimeout(() => resolve("timeout"), WORKER_CLOSE_GRACE_MS);
+      timeout.unref?.();
+    });
+    try {
+      const result = await Promise.race([closeRequest, closeTimeout]);
+      if (result === "timeout") {
+        closeRequest.catch(() => {});
+      }
+    } finally {
+      if (timeout) {
+        clearTimeout(timeout);
+      }
+      this.shutdownChild();
+    }
+  }
+
+  /** Ensure the child process exists and has lifecycle failure handlers installed. */
+  private ensureChild(): ChildProcess {
+    if (this.child?.connected) {
+      return this.child;
+    }
+
+    const child = fork(this.scriptPath, [], {
+      execArgv: resolveWorkerExecArgv(),
+      serialization: "json",
+      stdio: ["ignore", "ignore", "ignore", "ipc"],
+    });
+    child.on("message", (message) => this.handleMessage(message));
+    child.on("exit", (code, signal) => {
+      if (this.child === child) {
+        this.child = null;
+      }
+      this.rejectPending(createWorkerExitError(code, signal));
+    });
+    child.on("error", (err) => {
+      if (this.child === child) {
+        this.child = null;
+      }
+      this.rejectPending(
+        createLocalEmbeddingWorkerFailureError({
+          message: `Local embedding worker process failed: ${err.message}`,
+          code: LOCAL_EMBEDDING_WORKER_ERROR_CODES.processError,
+          reason: "process-error",
+          cause: err,
+        }),
+      );
+    });
+    this.child = child;
+    return child;
+  }
+
+  /** Send one request over IPC and bind its abort signal to child shutdown. */
+  private async send(
+    request: LocalEmbeddingWorkerRequestPayload,
+    options?: EmbeddingProviderCallOptions,
+  ): Promise<number[] | number[][] | undefined> {
+    options?.signal?.throwIfAborted();
+    const child = this.ensureChild();
+    const id = this.nextRequestId++;
+    const payload = { ...request, id } as LocalEmbeddingWorkerRequest;
+    return await new Promise((resolve, reject) => {
+      const pending: PendingRequest = { resolve, reject };
+      if (options?.signal) {
+        const abort = () => {
+          this.pending.delete(id);
+          this.shutdownChild();
+          reject(
+            toLintErrorObject(
+              options.signal?.reason ?? new Error("Local embedding request aborted"),
+              "Non-Error rejection",
+            ),
+          );
+        };
+        options.signal.addEventListener("abort", abort, { once: true });
+        pending.abort = () => options.signal?.removeEventListener("abort", abort);
+      }
+      this.pending.set(id, pending);
+      child.send(payload, (err) => {
+        if (err) {
+          this.pending.delete(id);
+          pending.abort?.();
+          reject(
+            createLocalEmbeddingWorkerFailureError({
+              message: `Local embedding worker IPC failed: ${err.message}`,
+              code: LOCAL_EMBEDDING_WORKER_ERROR_CODES.ipcError,
+              reason: "ipc",
+              cause: err,
+            }),
+          );
+        }
+      });
+    });
+  }
+
+  /** Route one worker response to the matching pending request. */
+  private handleMessage(message: unknown): void {
+    const response = message as Partial<LocalEmbeddingWorkerResponse>;
+    if (typeof response.id !== "number") {
+      return;
+    }
+    const pending = this.pending.get(response.id);
+    if (!pending) {
+      return;
+    }
+    this.pending.delete(response.id);
+    pending.abort?.();
+    if (response.ok) {
+      pending.resolve(response.value);
+      return;
+    }
+    pending.reject(
+      createWorkerResponseError(response as LocalEmbeddingWorkerResponse & { ok: false }),
+    );
+  }
+
+  /** Disconnect and kill the current child process if it is still alive. */
+  private shutdownChild(): void {
+    const child = this.child;
+    this.child = null;
+    if (!child) {
+      return;
+    }
+    if (child.connected) {
+      child.disconnect();
+    }
+    if (!child.killed) {
+      child.kill();
+    }
+  }
+
+  /** Reject all pending requests after child process failure. */
+  private rejectPending(err: unknown): void {
+    const pending = [...this.pending.values()];
+    this.pending.clear();
+    for (const entry of pending) {
+      entry.abort?.();
+      entry.reject(err);
+    }
+  }
+}
+
+/** Create the public local embedding provider backed by the child worker client. */
+export async function createLocalEmbeddingWorkerProvider(
+  options: EmbeddingProviderOptions,
+  runtimeOptions?: LocalEmbeddingProviderRuntimeOptions,
+): Promise<EmbeddingProvider> {
+  const modelPath = normalizeOptionalString(options.local?.modelPath) || DEFAULT_LOCAL_MODEL;
+  const workerOptions = serializeLocalEmbeddingOptions(options, runtimeOptions);
+  const client = new LocalEmbeddingWorkerClient(
+    runtimeOptions?.workerScriptPath ?? resolveDefaultWorkerScriptPath(),
+  );
+  try {
+    await client.initialize(workerOptions);
+  } catch (err) {
+    await client.close().catch(() => {});
+    throw err;
+  }
+  let closed = false;
+
+  const throwIfClosed = () => {
+    if (closed) {
+      throw new Error("Local embedding provider has been closed");
+    }
+  };
+
+  return {
+    id: "local",
+    model: modelPath,
+    embedQuery: async (text, callOptions) => {
+      throwIfClosed();
+      return await client.embedQuery(workerOptions, text, callOptions);
+    },
+    embedBatch: async (texts, callOptions) => {
+      throwIfClosed();
+      return await client.embedBatch(workerOptions, texts, callOptions);
+    },
+    close: async () => {
+      if (closed) {
+        return;
+      }
+      closed = true;
+      await client.close();
+    },
+  };
+}
+
+/** Convert abort reasons or arbitrary thrown values into lint-safe Error objects. */
+function toLintErrorObject(value: unknown, fallbackMessage: string): Error {
+  if (value instanceof Error) {
+    return value;
+  }
+  if (typeof value === "string") {
+    return new Error(value);
+  }
+  const error = new Error(fallbackMessage, { cause: value });
+  if ((typeof value === "object" && value !== null) || typeof value === "function") {
+    Object.assign(error, value);
+  }
+  return error;
+}

packages/memory-host-sdk/src/host/embeddings.test.ts

@@ -0,0 +1,625 @@
+// Memory Host SDK tests cover embeddings behavior.
+import fs from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+import { LOCAL_EMBEDDING_WORKER_ERROR_CODES } from "./embedding-worker-errors.js";
+import { createLocalEmbeddingWorkerProvider } from "./embeddings-worker.js";
+import { createLocalEmbeddingProviderInProcess, DEFAULT_LOCAL_MODEL } from "./embeddings.js";
+
+const nodeLlamaMock = vi.hoisted(() => ({
+  importNodeLlamaCpp: vi.fn(),
+}));
+
+vi.mock("./node-llama.js", () => ({
+  importNodeLlamaCpp: nodeLlamaMock.importNodeLlamaCpp,
+}));
+
+beforeEach(() => {
+  nodeLlamaMock.importNodeLlamaCpp.mockReset();
+});
+
+afterEach(() => {
+  vi.resetAllMocks();
+});
+
+function createDeferred<T>() {
+  let resolve: ((value: T) => void) | undefined;
+  let reject: ((reason?: unknown) => void) | undefined;
+  const promise = new Promise<T>((res, rej) => {
+    resolve = res;
+    reject = rej;
+  });
+  if (!resolve || !reject) {
+    throw new Error("Expected deferred callbacks to be initialized");
+  }
+  return { promise, resolve, reject };
+}
+
+function mockLocalEmbeddingRuntime(
+  vector: ArrayLike<number> = new Float32Array([2.35, 3.45, 0.63, 4.3]),
+) {
+  const disposeContext = vi.fn();
+  const disposeModel = vi.fn();
+  const disposeLlama = vi.fn();
+  const getEmbeddingFor = vi.fn().mockResolvedValue({ vector });
+  const createEmbeddingContext = vi
+    .fn()
+    .mockResolvedValue({ getEmbeddingFor, dispose: disposeContext });
+  const loadModel = vi.fn().mockResolvedValue({ createEmbeddingContext, dispose: disposeModel });
+  const getLlama = vi.fn(async () => ({ loadModel, dispose: disposeLlama }));
+  const resolveModelFile = vi.fn(async (modelPath: string) => `/resolved/${modelPath}`);
+
+  nodeLlamaMock.importNodeLlamaCpp.mockResolvedValue({
+    getLlama,
+    resolveModelFile,
+    LlamaLogLevel: { error: 0 },
+  } as never);
+
+  return {
+    createEmbeddingContext,
+    disposeContext,
+    disposeLlama,
+    disposeModel,
+    getLlama,
+    getEmbeddingFor,
+    loadModel,
+    resolveModelFile,
+  };
+}
+
+describe("local embedding provider", () => {
+  it("normalizes local embeddings and resolves the default local model", async () => {
+    const runtime = mockLocalEmbeddingRuntime();
+
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+    });
+
+    const embedding = await provider.embedQuery("test query");
+    const magnitude = Math.sqrt(embedding.reduce((sum, value) => sum + value * value, 0));
+
+    expect(DEFAULT_LOCAL_MODEL).toBe(
+      "hf:ggml-org/embeddinggemma-300m-qat-q8_0-GGUF/embeddinggemma-300m-qat-Q8_0.gguf",
+    );
+    expect(magnitude).toBeCloseTo(1, 5);
+    expect(runtime.resolveModelFile).toHaveBeenCalledWith(
+      DEFAULT_LOCAL_MODEL,
+      expect.objectContaining({ signal: expect.any(AbortSignal) }),
+    );
+    expect(runtime.loadModel).toHaveBeenCalledWith(
+      expect.objectContaining({
+        modelPath: `/resolved/${DEFAULT_LOCAL_MODEL}`,
+        loadSignal: expect.any(AbortSignal),
+      }),
+    );
+    expect(runtime.getEmbeddingFor).toHaveBeenCalledWith("test query");
+  });
+
+  it("truncates local embeddings before normalizing them", async () => {
+    mockLocalEmbeddingRuntime(new Float32Array([3, 4, 12]));
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+      outputDimensionality: 2,
+    });
+
+    await expect(provider.embedQuery("test query")).resolves.toEqual([0.6, 0.8]);
+    await expect(provider.embedBatch(["test document"])).resolves.toEqual([[0.6, 0.8]]);
+  });
+
+  it("does not read local embedding coordinates past outputDimensionality", async () => {
+    mockLocalEmbeddingRuntime({
+      length: 3,
+      0: 3,
+      1: 4,
+      get 2(): number {
+        throw new Error("tail coordinate should not be read");
+      },
+    });
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+      outputDimensionality: 2,
+    });
+
+    await expect(provider.embedQuery("test query")).resolves.toEqual([0.6, 0.8]);
+  });
+
+  it("passes default contextSize (4096) to createEmbeddingContext when not configured", async () => {
+    const runtime = mockLocalEmbeddingRuntime();
+
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+    });
+
+    await provider.embedQuery("context size default test");
+
+    expect(runtime.createEmbeddingContext).toHaveBeenCalledWith(
+      expect.objectContaining({ contextSize: 4096, createSignal: expect.any(AbortSignal) }),
+    );
+  });
+
+  it("imports node-llama-cpp from an explicit module URL when provided", async () => {
+    mockLocalEmbeddingRuntime();
+
+    await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+      local: {
+        nodeLlamaCppImportUrl: "file:///plugins/llama-cpp/node-llama-cpp.js",
+      } as never,
+    });
+
+    expect(nodeLlamaMock.importNodeLlamaCpp).toHaveBeenCalledWith(
+      "file:///plugins/llama-cpp/node-llama-cpp.js",
+    );
+  });
+
+  it("passes configured contextSize to createEmbeddingContext", async () => {
+    const runtime = mockLocalEmbeddingRuntime();
+
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+      local: { contextSize: 2048 },
+    });
+
+    await provider.embedQuery("context size custom test");
+
+    expect(runtime.createEmbeddingContext).toHaveBeenCalledWith(
+      expect.objectContaining({ contextSize: 2048, createSignal: expect.any(AbortSignal) }),
+    );
+  });
+
+  it('passes "auto" contextSize to createEmbeddingContext when explicitly set', async () => {
+    const runtime = mockLocalEmbeddingRuntime();
+
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+      local: { contextSize: "auto" },
+    });
+
+    await provider.embedQuery("context size auto test");
+
+    expect(runtime.createEmbeddingContext).toHaveBeenCalledWith(
+      expect.objectContaining({ contextSize: "auto", createSignal: expect.any(AbortSignal) }),
+    );
+  });
+
+  it("runs local batch embeddings sequentially", async () => {
+    const calls: string[] = [];
+    const firstGate = createDeferred<{ vector: Float32Array }>();
+    const secondGate = createDeferred<{ vector: Float32Array }>();
+    const getEmbeddingFor = vi.fn((text: string) => {
+      calls.push(text);
+      return text === "first" ? firstGate.promise : secondGate.promise;
+    });
+    nodeLlamaMock.importNodeLlamaCpp.mockResolvedValue({
+      getLlama: vi.fn(async () => ({
+        loadModel: vi.fn(async () => ({
+          createEmbeddingContext: vi.fn(async () => ({ getEmbeddingFor })),
+        })),
+      })),
+      resolveModelFile: vi.fn(async () => "/resolved/model.gguf"),
+      LlamaLogLevel: { error: 0 },
+    } as never);
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+    });
+
+    const batchPromise = provider.embedBatch(["first", "second"]);
+    await expect.poll(() => calls.join(",")).toBe("first");
+    firstGate.resolve({ vector: new Float32Array([1, 0]) });
+    await expect.poll(() => calls.join(",")).toBe("first,second");
+    secondGate.resolve({ vector: new Float32Array([0, 1]) });
+
+    await expect(batchPromise).resolves.toHaveLength(2);
+  });
+
+  it("trims explicit local model paths and cache directories", async () => {
+    const runtime = mockLocalEmbeddingRuntime(new Float32Array([1, 0]));
+
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+      local: {
+        modelPath: "  /models/embed.gguf  ",
+        modelCacheDir: "  /cache/models  ",
+      },
+    });
+
+    await provider.embedBatch(["a", "b"]);
+
+    expect(provider.model).toBe("/models/embed.gguf");
+    expect(runtime.resolveModelFile).toHaveBeenCalledWith(
+      "/models/embed.gguf",
+      expect.objectContaining({
+        directory: "/cache/models",
+        signal: expect.any(AbortSignal),
+      }),
+    );
+    expect(runtime.getEmbeddingFor).toHaveBeenCalledTimes(2);
+  });
+
+  it("disposes cached local llama resources when closed", async () => {
+    const runtime = mockLocalEmbeddingRuntime();
+
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+    });
+
+    await provider.embedQuery("load local resources");
+    await provider.close?.();
+    await provider.close?.();
+
+    expect(runtime.disposeContext).toHaveBeenCalledTimes(1);
+    expect(runtime.disposeModel).toHaveBeenCalledTimes(1);
+    expect(runtime.disposeLlama).toHaveBeenCalledTimes(1);
+    await expect(provider.embedQuery("after close")).rejects.toThrow(
+      "Local embedding provider has been closed",
+    );
+  });
+
+  it("does not wait for pending local llama initialization before close resolves", async () => {
+    const disposeLlama = vi.fn();
+    const getLlamaGate = createDeferred<unknown>();
+    nodeLlamaMock.importNodeLlamaCpp.mockResolvedValue({
+      getLlama: async () => (await getLlamaGate.promise) as never,
+      resolveModelFile: vi.fn(async (modelPath: string) => `/resolved/${modelPath}`),
+      LlamaLogLevel: { error: 0 },
+    } as never);
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+    });
+
+    const embedPromise = provider.embedQuery("pending init");
+    await expect(provider.close?.()).resolves.toBeUndefined();
+
+    getLlamaGate.resolve({ loadModel: vi.fn(), dispose: disposeLlama });
+    await expect(embedPromise).rejects.toThrow("Local embedding provider has been closed");
+    expect(disposeLlama).toHaveBeenCalledTimes(1);
+  });
+
+  it("aborts pending local llama model loads when closed", async () => {
+    const loadModelStarted = createDeferred<void>();
+    const loadModelGate = createDeferred<never>();
+    const disposeLlama = vi.fn();
+    let capturedResolveSignal: AbortSignal | undefined;
+    let capturedLoadSignal: AbortSignal | undefined;
+    const loadModel = vi.fn(
+      (params: { modelPath: string; loadSignal?: AbortSignal }): Promise<never> => {
+        capturedLoadSignal = params.loadSignal;
+        loadModelStarted.resolve();
+        return loadModelGate.promise;
+      },
+    );
+    nodeLlamaMock.importNodeLlamaCpp.mockResolvedValue({
+      getLlama: async () => ({ loadModel, dispose: disposeLlama }),
+      resolveModelFile: vi.fn(async (_modelPath: string, options?: { signal?: AbortSignal }) => {
+        capturedResolveSignal = options?.signal;
+        return "/resolved/model.gguf";
+      }),
+      LlamaLogLevel: { error: 0 },
+    } as never);
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+    });
+
+    const embedPromise = provider.embedQuery("pending model load");
+    await loadModelStarted.promise;
+    await expect(provider.close?.()).resolves.toBeUndefined();
+
+    expect(capturedResolveSignal?.aborted).toBe(true);
+    expect(capturedLoadSignal?.aborted).toBe(true);
+    expect(disposeLlama).toHaveBeenCalledTimes(1);
+    loadModelGate.reject(new Error("load aborted"));
+    await expect(embedPromise).rejects.toThrow("load aborted");
+  });
+
+  it("aborts pending local llama embedding context creation when closed", async () => {
+    const createContextStarted = createDeferred<void>();
+    const createContextGate = createDeferred<never>();
+    const disposeLlama = vi.fn();
+    const disposeModel = vi.fn();
+    let capturedCreateSignal: AbortSignal | undefined;
+    const createEmbeddingContext = vi.fn(
+      (options?: { createSignal?: AbortSignal }): Promise<never> => {
+        capturedCreateSignal = options?.createSignal;
+        createContextStarted.resolve();
+        return createContextGate.promise;
+      },
+    );
+    nodeLlamaMock.importNodeLlamaCpp.mockResolvedValue({
+      getLlama: async () => ({
+        loadModel: vi.fn(async () => ({ createEmbeddingContext, dispose: disposeModel })),
+        dispose: disposeLlama,
+      }),
+      resolveModelFile: vi.fn(async () => "/resolved/model.gguf"),
+      LlamaLogLevel: { error: 0 },
+    } as never);
+    const provider = await createLocalEmbeddingProviderInProcess({
+      config: {} as never,
+      provider: "local",
+      model: "",
+      fallback: "none",
+    });
+
+    const embedPromise = provider.embedQuery("pending context create");
+    await createContextStarted.promise;
+    await expect(provider.close?.()).resolves.toBeUndefined();
+
+    expect(capturedCreateSignal?.aborted).toBe(true);
+    expect(disposeModel).toHaveBeenCalledTimes(1);
+    expect(disposeLlama).toHaveBeenCalledTimes(1);
+    createContextGate.reject(new Error("context create aborted"));
+    await expect(embedPromise).rejects.toThrow("context create aborted");
+  });
+
+  it("uses a worker process for the public local provider", async () => {
+    const tempDir = await fs.mkdtemp(path.join(os.tmpdir(), "openclaw-local-embedding-worker-"));
+    const workerScript = path.join(tempDir, "worker.cjs");
+    await fs.writeFile(
+      workerScript,
+      `
+process.on("message", (message) => {
+  if (message.type === "initialize") {
+    if (message.options.local?.nodeLlamaCppImportUrl !== "file:///plugin/node-llama-cpp.js") {
+      process.send({ id: message.id, ok: false, error: "missing nodeLlamaCppImportUrl" });
+      return;
+    }
+    if (message.options.outputDimensionality !== 2) {
+      process.send({ id: message.id, ok: false, error: "missing outputDimensionality" });
+      return;
+    }
+    process.send({ id: message.id, ok: true });
+    return;
+  }
+  if (message.type === "embedQuery") {
+    process.send({ id: message.id, ok: true, value: [1, 0] });
+    return;
+  }
+  if (message.type === "embedBatch") {
+    process.send({ id: message.id, ok: true, value: message.texts.map(() => [0, 1]) });
+    return;
+  }
+  process.send({ id: message.id, ok: true });
+});
+`,
+      "utf8",
+    );
+    const provider = await createLocalEmbeddingWorkerProvider(
+      {
+        config: {} as never,
+        provider: "local",
+        model: "",
+        fallback: "none",
+        outputDimensionality: 2,
+      },
+      {
+        workerScriptPath: workerScript,
+        nodeLlamaCppImportUrl: "file:///plugin/node-llama-cpp.js",
+      },
+    );
+
+    await expect(provider.embedQuery("hello")).resolves.toEqual([1, 0]);
+    await expect(provider.embedBatch(["a", "b"])).resolves.toEqual([
+      [0, 1],
+      [0, 1],
+    ]);
+    await expect(provider.close?.()).resolves.toBeUndefined();
+  });
+
+  it("terminates the worker when close runs behind a pending request", async () => {
+    const tempDir = await fs.mkdtemp(path.join(os.tmpdir(), "openclaw-local-embedding-worker-"));
+    const workerScript = path.join(tempDir, "worker.cjs");
+    const embedStartedPath = path.join(tempDir, "embed-started");
+    await fs.writeFile(
+      workerScript,
+      `
+const fs = require("node:fs");
+const embedStartedPath = ${JSON.stringify(embedStartedPath)};
+let busy = false;
+
+process.on("message", (message) => {
+  if (busy) {
+    return;
+  }
+  if (message.type === "initialize") {
+    process.send({ id: message.id, ok: true });
+    return;
+  }
+  if (message.type === "embedQuery") {
+    busy = true;
+    fs.writeFileSync(embedStartedPath, "1");
+  }
+});
+`,
+      "utf8",
+    );
+    const provider = await createLocalEmbeddingWorkerProvider(
+      {
+        config: {} as never,
+        provider: "local",
+        model: "",
+        fallback: "none",
+      },
+      { workerScriptPath: workerScript },
+    );
+
+    const embedPromise = provider.embedQuery("stuck");
+    const embedError = embedPromise.then(
+      () => undefined,
+      (err: unknown) => err,
+    );
+    await expect
+      .poll(async () => {
+        try {
+          await fs.access(embedStartedPath);
+          return true;
+        } catch {
+          return false;
+        }
+      })
+      .toBe(true);
+
+    const closePromise = provider.close?.() ?? Promise.resolve();
+    const closeResult = await Promise.race([
+      closePromise.then(() => "closed" as const),
+      new Promise<"timeout">((resolve) => {
+        setTimeout(() => resolve("timeout"), 1_000);
+      }),
+    ]);
+
+    expect(closeResult).toBe("closed");
+    await expect(embedError).resolves.toMatchObject({
+      code: LOCAL_EMBEDDING_WORKER_ERROR_CODES.exited,
+    });
+  });
+
+  it("does not pass inline-source or inspector exec args to the file-backed worker", async () => {
+    const tempDir = await fs.mkdtemp(path.join(os.tmpdir(), "openclaw-local-embedding-worker-"));
+    const workerScript = path.join(tempDir, "worker.cjs");
+    await fs.writeFile(
+      workerScript,
+      `
+process.on("message", (message) => {
+  if (message.type === "initialize" || message.type === "close") {
+    process.send({ id: message.id, ok: true });
+    return;
+  }
+  process.send({ id: message.id, ok: true, value: [process.execArgv.length] });
+});
+`,
+      "utf8",
+    );
+    const originalExecArgv = [...process.execArgv];
+    let provider: Awaited<ReturnType<typeof createLocalEmbeddingWorkerProvider>> | undefined;
+    try {
+      process.execArgv.splice(
+        0,
+        process.execArgv.length,
+        "--eval",
+        "setInterval(() => {}, 1000)",
+        "--print",
+        "1 + 1",
+        "--input-type=module",
+        "--inspect-brk=127.0.0.1:0",
+        "--inspect-port",
+        "0",
+      );
+      provider = await createLocalEmbeddingWorkerProvider(
+        {
+          config: {} as never,
+          provider: "local",
+          model: "",
+          fallback: "none",
+        },
+        { workerScriptPath: workerScript },
+      );
+      await expect(provider.embedQuery("hello")).resolves.toEqual([0]);
+    } finally {
+      process.execArgv.splice(0, process.execArgv.length, ...originalExecArgv);
+      await provider?.close?.();
+    }
+  });
+
+  it("reports worker initialization failures during provider creation", async () => {
+    const tempDir = await fs.mkdtemp(path.join(os.tmpdir(), "openclaw-local-embedding-worker-"));
+    const workerScript = path.join(tempDir, "worker.cjs");
+    await fs.writeFile(
+      workerScript,
+      `
+process.on("message", (message) => {
+  process.send({
+    id: message.id,
+    ok: false,
+    error: { message: "Cannot find package 'node-llama-cpp'", code: "ERR_MODULE_NOT_FOUND" },
+  });
+});
+`,
+      "utf8",
+    );
+
+    try {
+      await createLocalEmbeddingWorkerProvider(
+        {
+          config: {} as never,
+          provider: "local",
+          model: "",
+          fallback: "none",
+        },
+        { workerScriptPath: workerScript },
+      );
+      throw new Error("expected local embedding provider creation to fail");
+    } catch (err) {
+      expect(err).toBeInstanceOf(Error);
+      expect((err as Error).message).toBe("Cannot find package 'node-llama-cpp'");
+      expect((err as Error & { code?: string }).code).toBe("ERR_MODULE_NOT_FOUND");
+    }
+  });
+
+  it("reports worker exits with structured failure codes", async () => {
+    const tempDir = await fs.mkdtemp(path.join(os.tmpdir(), "openclaw-local-embedding-worker-"));
+    const workerScript = path.join(tempDir, "worker.cjs");
+    await fs.writeFile(
+      workerScript,
+      `
+process.on("message", (message) => {
+  if (message.type === "initialize") {
+    process.send({ id: message.id, ok: true });
+    return;
+  }
+  process.exit(134);
+});
+`,
+      "utf8",
+    );
+    const provider = await createLocalEmbeddingWorkerProvider(
+      {
+        config: {} as never,
+        provider: "local",
+        model: "",
+        fallback: "none",
+      },
+      { workerScriptPath: workerScript },
+    );
+
+    await expect(provider.embedQuery("hello")).rejects.toMatchObject({
+      code: LOCAL_EMBEDDING_WORKER_ERROR_CODES.exited,
+      reason: "exit",
+      exitCode: 134,
+    });
+  });
+});

packages/memory-host-sdk/src/host/embeddings.ts

@@ -0,0 +1,206 @@
+import { DEFAULT_LOCAL_MODEL } from "./embedding-defaults.js";
+import { sanitizeAndNormalizeEmbedding } from "./embedding-vectors.js";
+import { createLocalEmbeddingWorkerProvider } from "./embeddings-worker.js";
+import type { EmbeddingProvider, EmbeddingProviderOptions } from "./embeddings.types.js";
+import {
+  importNodeLlamaCpp,
+  type Llama,
+  type LlamaEmbeddingContext,
+  type LlamaModel,
+} from "./node-llama.js";
+// Memory Host SDK module implements embeddings behavior.
+import { toLintErrorObject } from "./retry-utils.js";
+import { normalizeOptionalString } from "./string-utils.js";
+
+type DisposableResource = {
+  dispose?: () => Promise<void> | void;
+};
+
+export type {
+  EmbeddingProvider,
+  EmbeddingProviderFallback,
+  EmbeddingProviderId,
+  EmbeddingProviderOptions,
+  EmbeddingProviderRequest,
+  GeminiTaskType,
+} from "./embeddings.types.js";
+
+export { DEFAULT_LOCAL_MODEL } from "./embedding-defaults.js";
+
+export type LocalEmbeddingProviderRuntimeOptions = {
+  workerScriptPath?: string;
+  nodeLlamaCppImportUrl?: string;
+};
+
+function copyEmbeddingVector(vector: ArrayLike<number>, maxLength?: number): number[] {
+  const length = Math.min(maxLength ?? vector.length, vector.length);
+  const values: number[] = [];
+  for (let index = 0; index < length; index += 1) {
+    values.push(vector[index]);
+  }
+  return values;
+}
+
+async function disposeResources(
+  resources: Array<DisposableResource | null | undefined>,
+): Promise<void> {
+  let firstError: unknown;
+  for (const resource of resources) {
+    try {
+      await resource?.dispose?.();
+    } catch (err) {
+      firstError ??= err;
+    }
+  }
+  if (firstError) {
+    throw toLintErrorObject(firstError, "Non-Error thrown");
+  }
+}
+
+export async function createLocalEmbeddingProvider(
+  options: EmbeddingProviderOptions,
+  runtimeOptions?: LocalEmbeddingProviderRuntimeOptions,
+): Promise<EmbeddingProvider> {
+  return await createLocalEmbeddingWorkerProvider(options, runtimeOptions);
+}
+
+export async function createLocalEmbeddingProviderInProcess(
+  options: EmbeddingProviderOptions,
+): Promise<EmbeddingProvider> {
+  const modelPath = normalizeOptionalString(options.local?.modelPath) || DEFAULT_LOCAL_MODEL;
+  const modelCacheDir = normalizeOptionalString(options.local?.modelCacheDir);
+  const nodeLlamaCppImportUrl = normalizeOptionalString(
+    (options.local as EmbeddingProviderOptions["local"] & { nodeLlamaCppImportUrl?: string })
+      ?.nodeLlamaCppImportUrl,
+  );
+  const contextSize: number | "auto" = options.local?.contextSize ?? 4096;
+
+  // Lazy-load node-llama-cpp to keep startup light unless local is enabled.
+  const { getLlama, resolveModelFile, LlamaLogLevel } =
+    await importNodeLlamaCpp(nodeLlamaCppImportUrl);
+
+  let llama: Llama | null = null;
+  let embeddingModel: LlamaModel | null = null;
+  let embeddingContext: LlamaEmbeddingContext | null = null;
+  let initPromise: Promise<LlamaEmbeddingContext> | null = null;
+  let initAbortController: AbortController | null = null;
+  let closePromise: Promise<void> | null = null;
+  let closed = false;
+
+  const throwIfClosed = () => {
+    if (closed) {
+      throw new Error("Local embedding provider has been closed");
+    }
+  };
+  const disposeAndThrowIfClosed = async <T extends DisposableResource>(resource: T): Promise<T> => {
+    if (!closed) {
+      return resource;
+    }
+    await disposeResources([resource]);
+    throwIfClosed();
+    return resource;
+  };
+
+  const ensureContext = async (): Promise<LlamaEmbeddingContext> => {
+    throwIfClosed();
+    if (embeddingContext) {
+      return embeddingContext;
+    }
+    if (initPromise) {
+      return initPromise;
+    }
+    initPromise = (async () => {
+      const abortController = new AbortController();
+      initAbortController = abortController;
+      try {
+        if (!llama) {
+          const nextLlama = await getLlama({
+            logLevel: LlamaLogLevel.error,
+          });
+          llama = await disposeAndThrowIfClosed(nextLlama);
+        }
+        if (!embeddingModel) {
+          const resolved = await resolveModelFile(modelPath, {
+            ...(modelCacheDir ? { directory: modelCacheDir } : {}),
+            signal: abortController.signal,
+          });
+          throwIfClosed();
+          const nextModel = await llama.loadModel({
+            modelPath: resolved,
+            loadSignal: abortController.signal,
+          });
+          embeddingModel = await disposeAndThrowIfClosed(nextModel);
+        }
+        if (!embeddingContext) {
+          const nextContext = await embeddingModel.createEmbeddingContext({
+            contextSize,
+            createSignal: abortController.signal,
+          });
+          embeddingContext = await disposeAndThrowIfClosed(nextContext);
+        }
+        return embeddingContext;
+      } catch (err) {
+        initPromise = null;
+        throw err;
+      } finally {
+        if (initAbortController === abortController) {
+          initAbortController = null;
+        }
+      }
+    })();
+    return initPromise;
+  };
+
+  const outputDimensionality =
+    typeof options.outputDimensionality === "number" ? options.outputDimensionality : undefined;
+  const normalize = (vector: ArrayLike<number>): number[] =>
+    sanitizeAndNormalizeEmbedding(copyEmbeddingVector(vector, outputDimensionality));
+
+  return {
+    id: "local",
+    model: modelPath,
+    embedQuery: async (text, optionsValue) => {
+      throwIfClosed();
+      optionsValue?.signal?.throwIfAborted();
+      const ctx = await ensureContext();
+      throwIfClosed();
+      optionsValue?.signal?.throwIfAborted();
+      const embedding = await ctx.getEmbeddingFor(text);
+      return normalize(embedding.vector);
+    },
+    embedBatch: async (texts, optionsLocal) => {
+      throwIfClosed();
+      optionsLocal?.signal?.throwIfAborted();
+      const ctx = await ensureContext();
+      throwIfClosed();
+      optionsLocal?.signal?.throwIfAborted();
+      const embeddings: number[][] = [];
+      for (const text of texts) {
+        throwIfClosed();
+        optionsLocal?.signal?.throwIfAborted();
+        const embedding = await ctx.getEmbeddingFor(text);
+        embeddings.push(normalize(embedding.vector));
+      }
+      return embeddings;
+    },
+    close: async () => {
+      if (closePromise) {
+        return closePromise;
+      }
+      closed = true;
+      initAbortController?.abort();
+      initAbortController = null;
+      closePromise = (async () => {
+        const context = embeddingContext;
+        const model = embeddingModel;
+        const runtime = llama;
+        embeddingContext = null;
+        embeddingModel = null;
+        llama = null;
+        initPromise = null;
+        await disposeResources([context, model, runtime]);
+      })();
+      return closePromise;
+    },
+  };
+}

packages/memory-host-sdk/src/host/embeddings.types.ts

@@ -0,0 +1,65 @@
+// Memory Host SDK type module defines shared TypeScript contracts.
+import type { OpenClawConfig, SecretInput } from "../engine-foundation.js";
+import type { EmbeddingInput } from "./embedding-inputs.js";
+
+export type EmbeddingProvider = {
+  id: string;
+  model: string;
+  maxInputTokens?: number;
+  embedQuery: (text: string, options?: EmbeddingProviderCallOptions) => Promise<number[]>;
+  embedBatch: (texts: string[], options?: EmbeddingProviderCallOptions) => Promise<number[][]>;
+  embedBatchInputs?: (
+    inputs: EmbeddingInput[],
+    options?: EmbeddingProviderCallOptions,
+  ) => Promise<number[][]>;
+  close?: () => Promise<void> | void;
+};
+
+export type EmbeddingProviderCallOptions = {
+  signal?: AbortSignal;
+};
+
+export type EmbeddingProviderId = string;
+export type EmbeddingProviderRequest = string;
+export type EmbeddingProviderFallback = string;
+
+export type GeminiTaskType =
+  | "RETRIEVAL_QUERY"
+  | "RETRIEVAL_DOCUMENT"
+  | "SEMANTIC_SIMILARITY"
+  | "CLASSIFICATION"
+  | "CLUSTERING"
+  | "QUESTION_ANSWERING"
+  | "FACT_VERIFICATION";
+
+export type EmbeddingProviderOptions = {
+  config: OpenClawConfig;
+  agentDir?: string;
+  provider?: EmbeddingProviderRequest;
+  remote?: {
+    baseUrl?: string;
+    apiKey?: SecretInput;
+    headers?: Record<string, string>;
+  };
+  model: string;
+  inputType?: string;
+  queryInputType?: string;
+  documentInputType?: string;
+  fallback?: EmbeddingProviderFallback;
+  local?: {
+    modelPath?: string;
+    modelCacheDir?: string;
+    /**
+     * Context size passed to node-llama-cpp `createEmbeddingContext`.
+     * Default: 4096, chosen to cover typical memory-search chunks (128–512 tokens)
+     * while keeping non-weight VRAM bounded.
+     * Set `"auto"` to let node-llama-cpp use the model's trained maximum — not
+     * recommended for 8B+ models (e.g. Qwen3-Embedding-8B: up to 40 960 tokens → ~32 GB VRAM).
+     */
+    contextSize?: number | "auto";
+  };
+  /** Provider-specific output vector dimensions for supported embedding families. */
+  outputDimensionality?: number;
+  /** Gemini: override the default task type sent with embedding requests. */
+  taskType?: GeminiTaskType;
+};

packages/memory-host-sdk/src/host/error-utils.ts

@@ -0,0 +1,92 @@
+// Memory Host SDK helper module supports error utils behavior.
+const SECRET_PATTERNS: RegExp[] = [
+  /\b[A-Z0-9_]*(?:KEY|TOKEN|SECRET|PASSWORD|PASSWD)\b\s*[=:]\s*(["']?)([^\s"'\\]+)\1/g,
+  /[?&](?:access[-_]?token|auth[-_]?token|hook[-_]?token|refresh[-_]?token|api[-_]?key|client[-_]?secret|token|key|secret|password|pass|passwd|auth|signature)=([^&\s"'<>]+)/gi,
+  /"(?:apiKey|token|secret|password|passwd|accessToken|refreshToken)"\s*:\s*"([^"]+)"/g,
+  /--(?:api[-_]?key|hook[-_]?token|token|secret|password|passwd)\s+(["']?)([^\s"']+)\1/g,
+  /Authorization\s*[:=]\s*Bearer\s+([A-Za-z0-9._\-+=]+)/g,
+  /\bBearer\s+([A-Za-z0-9._\-+=]{18,})\b/g,
+  /(^|[\s,;])(?:access_token|refresh_token|api[-_]?key|token|secret|password|passwd)=([^\s&#]+)/g,
+  /-----BEGIN [A-Z ]*PRIVATE KEY-----[\s\S]+?-----END [A-Z ]*PRIVATE KEY-----/g,
+  /\b(sk-[A-Za-z0-9_-]{8,})\b/g,
+  /\b(ghp_[A-Za-z0-9]{20,})\b/g,
+  /\b(github_pat_[A-Za-z0-9_]{20,})\b/g,
+  /\b(xox[baprs]-[A-Za-z0-9-]{10,})\b/g,
+  /\b(xapp-[A-Za-z0-9-]{10,})\b/g,
+  /\b(gsk_[A-Za-z0-9_-]{10,})\b/g,
+  /\b(AIza[0-9A-Za-z\-_]{20,})\b/g,
+  /\b(pplx-[A-Za-z0-9_-]{10,})\b/g,
+  /\b(npm_[A-Za-z0-9]{10,})\b/g,
+  /\bbot(\d{6,}:[A-Za-z0-9_-]{20,})\b/g,
+  /\b(\d{6,}:[A-Za-z0-9_-]{20,})\b/g,
+];
+
+// Redact common token/key shapes before errors leave memory host internals.
+function maskToken(token: string): string {
+  if (token.length < 18) {
+    return "***";
+  }
+  return `${token.slice(0, 6)}...${token.slice(-4)}`;
+}
+
+function redactPemBlock(block: string): string {
+  const lines = block.split(/\r?\n/).filter(Boolean);
+  if (lines.length < 2) {
+    return "***";
+  }
+  return `${lines[0]}\n...redacted...\n${lines[lines.length - 1]}`;
+}
+
+function redactMatch(match: string, groups: string[]): string {
+  if (match.includes("PRIVATE KEY-----")) {
+    return redactPemBlock(match);
+  }
+  const token = groups.findLast((value) => typeof value === "string" && value.length > 0) ?? match;
+  const masked = maskToken(token);
+  return token === match ? masked : match.replace(token, masked);
+}
+
+function redactSensitiveText(text: string): string {
+  let next = text;
+  for (const pattern of SECRET_PATTERNS) {
+    next = next.replace(pattern, (...args: string[]) =>
+      redactMatch(args[0] ?? "", args.slice(1, -2)),
+    );
+  }
+  return next;
+}
+
+/** Format unknown errors with causes while redacting likely secrets. */
+export function formatErrorMessage(err: unknown): string {
+  let formatted: string;
+  if (err instanceof Error) {
+    formatted = err.message || err.name || "Error";
+    let cause: unknown = err.cause;
+    const seen = new Set<unknown>([err]);
+    while (cause && !seen.has(cause)) {
+      seen.add(cause);
+      if (cause instanceof Error) {
+        if (cause.message) {
+          formatted += ` | ${cause.message}`;
+        }
+        cause = cause.cause;
+      } else if (typeof cause === "string") {
+        formatted += ` | ${cause}`;
+        break;
+      } else {
+        break;
+      }
+    }
+  } else if (typeof err === "string") {
+    formatted = err;
+  } else if (typeof err === "number" || typeof err === "boolean" || typeof err === "bigint") {
+    formatted = String(err);
+  } else {
+    try {
+      formatted = JSON.stringify(err);
+    } catch {
+      formatted = Object.prototype.toString.call(err);
+    }
+  }
+  return redactSensitiveText(formatted);
+}

packages/memory-host-sdk/src/host/fs-utils.ts

@@ -0,0 +1,34 @@
+// Memory Host SDK helper module supports fs utils behavior.
+import { configureFsSafePython } from "@openclaw/fs-safe/config";
+// fs-safe facade with Python validation disabled by default for this package's
+// host-side memory file operations.
+export { root } from "@openclaw/fs-safe/root";
+export { isPathInside, isPathInsideWithRealpath } from "@openclaw/fs-safe/path";
+export {
+  assertNoSymlinkParents,
+  readRegularFile,
+  statRegularFile,
+  type RegularFileStatResult,
+} from "@openclaw/fs-safe/advanced";
+export { walkDirectory, type WalkDirectoryEntry } from "@openclaw/fs-safe/walk";
+
+const hasPythonModeOverride =
+  process.env.FS_SAFE_PYTHON_MODE != null || process.env.OPENCLAW_FS_SAFE_PYTHON_MODE != null;
+
+if (!hasPythonModeOverride) {
+  configureFsSafePython({ mode: "off" });
+}
+
+/** True for missing-file errors emitted by Node or fs-safe. */
+export function isFileMissingError(
+  err: unknown,
+): err is NodeJS.ErrnoException & { code: "ENOENT" | "ENOTDIR" | "not-found" } {
+  return Boolean(
+    err &&
+    typeof err === "object" &&
+    "code" in err &&
+    ((err as Partial<NodeJS.ErrnoException>).code === "ENOENT" ||
+      (err as Partial<NodeJS.ErrnoException>).code === "ENOTDIR" ||
+      (err as { code?: unknown }).code === "not-found"),
+  );
+}

packages/memory-host-sdk/src/host/hash.ts

@@ -0,0 +1,7 @@
+// Memory Host SDK module implements hash behavior.
+import crypto from "node:crypto";
+
+/** SHA-256 hash helper for stable cache/content keys. */
+export function hashText(value: string): string {
+  return crypto.createHash("sha256").update(value).digest("hex");
+}

packages/memory-host-sdk/src/host/internal.test.ts

@@ -0,0 +1,250 @@
+// Memory Host SDK tests cover internal behavior.
+import fsSync from "node:fs";
+import fs from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { afterAll, afterEach, beforeAll, beforeEach, describe, expect, it, vi } from "vitest";
+import {
+  buildFileEntry,
+  buildMultimodalChunkForIndexing,
+  chunkMarkdown,
+  ensureDir,
+  isMemoryPath,
+  listMemoryFiles,
+  normalizeExtraMemoryPaths,
+  remapChunkLines,
+} from "./internal.js";
+import {
+  DEFAULT_MEMORY_MULTIMODAL_MAX_FILE_BYTES,
+  type MemoryMultimodalSettings,
+} from "./multimodal.js";
+
+type FileEntry = NonNullable<Awaited<ReturnType<typeof buildFileEntry>>>;
+type MultimodalIndexingChunk = NonNullable<
+  Awaited<ReturnType<typeof buildMultimodalChunkForIndexing>>
+>;
+
+let sharedTempRoot = "";
+let sharedTempId = 0;
+
+beforeAll(() => {
+  sharedTempRoot = fsSync.mkdtempSync(path.join(os.tmpdir(), "memory-host-sdk-package-tests-"));
+});
+
+afterAll(() => {
+  if (sharedTempRoot) {
+    fsSync.rmSync(sharedTempRoot, { recursive: true, force: true });
+  }
+});
+
+afterEach(() => {
+  vi.restoreAllMocks();
+});
+
+function setupTempDirLifecycle(prefix: string): () => string {
+  let tmpDir = "";
+  beforeEach(() => {
+    tmpDir = path.join(sharedTempRoot, `${prefix}${sharedTempId++}`);
+    fsSync.mkdirSync(tmpDir, { recursive: true });
+  });
+  return () => tmpDir;
+}
+
+function expectFileEntry(entry: Awaited<ReturnType<typeof buildFileEntry>>): FileEntry {
+  if (!entry) {
+    throw new Error("Expected file entry to be built");
+  }
+  return entry;
+}
+
+function expectMultimodalIndexingChunk(
+  built: Awaited<ReturnType<typeof buildMultimodalChunkForIndexing>>,
+): MultimodalIndexingChunk {
+  if (!built) {
+    throw new Error("Expected multimodal indexing chunk to be built");
+  }
+  return built;
+}
+
+function expectEmbeddingInput(
+  chunk: MultimodalIndexingChunk["chunk"],
+): NonNullable<MultimodalIndexingChunk["chunk"]["embeddingInput"]> {
+  if (!chunk.embeddingInput) {
+    throw new Error("Expected multimodal chunk embedding input");
+  }
+  return chunk.embeddingInput;
+}
+
+const multimodal: MemoryMultimodalSettings = {
+  enabled: true,
+  modalities: ["image", "audio"],
+  maxFileBytes: DEFAULT_MEMORY_MULTIMODAL_MAX_FILE_BYTES,
+};
+
+describe("memory host SDK package internals", () => {
+  const getTmpDir = setupTempDirLifecycle("memory-package-");
+
+  it("propagates directory creation failures", () => {
+    const mkdirError = new Error("disk full");
+    const targetDir = path.join(getTmpDir(), "blocked");
+    const mkdirSync = vi.spyOn(fsSync, "mkdirSync").mockImplementation(() => {
+      throw mkdirError;
+    });
+
+    expect(() => ensureDir(targetDir)).toThrow(mkdirError);
+    expect(mkdirSync).toHaveBeenCalledWith(targetDir, { recursive: true });
+  });
+
+  it("normalizes additional memory paths", () => {
+    const workspaceDir = path.join(os.tmpdir(), "memory-test-workspace");
+    const absPath = path.resolve(path.sep, "shared-notes");
+    expect(
+      normalizeExtraMemoryPaths(workspaceDir, [
+        " notes ",
+        "./notes",
+        absPath,
+        absPath,
+        "~/shared-notes",
+        "~",
+        "",
+      ]),
+    ).toEqual([
+      path.resolve(workspaceDir, "notes"),
+      absPath,
+      path.join(os.homedir(), "shared-notes"),
+      os.homedir(),
+    ]);
+  });
+
+  it("lists canonical markdown and enabled multimodal files", async () => {
+    const tmpDir = getTmpDir();
+    fsSync.writeFileSync(path.join(tmpDir, "MEMORY.md"), "# Default memory");
+    fsSync.writeFileSync(path.join(tmpDir, "memory.md"), "# Legacy memory");
+    const extraDir = path.join(tmpDir, "extra");
+    fsSync.mkdirSync(extraDir, { recursive: true });
+    fsSync.writeFileSync(path.join(extraDir, "note.md"), "# Note");
+    fsSync.writeFileSync(path.join(extraDir, "diagram.png"), Buffer.from("png"));
+    fsSync.writeFileSync(path.join(extraDir, "ignore.txt"), "ignored");
+
+    const files = await listMemoryFiles(
+      tmpDir,
+      [path.join(tmpDir, "memory.md"), extraDir],
+      multimodal,
+    );
+
+    expect(files.map((file) => path.relative(tmpDir, file)).toSorted()).toEqual([
+      "MEMORY.md",
+      path.join("extra", "diagram.png"),
+      path.join("extra", "note.md"),
+    ]);
+  });
+
+  it("allows top-level dreams path casing variants", () => {
+    expect(isMemoryPath("dreams.md")).toBe(true);
+    expect(isMemoryPath("DREAMS.md")).toBe(true);
+  });
+
+  it("builds markdown and multimodal file entries", async () => {
+    const tmpDir = getTmpDir();
+    const notePath = path.join(tmpDir, "note.md");
+    const imagePath = path.join(tmpDir, "diagram.png");
+    fsSync.writeFileSync(notePath, "hello", "utf-8");
+    fsSync.writeFileSync(imagePath, Buffer.from("png"));
+
+    const note = await buildFileEntry(notePath, tmpDir);
+    const image = await buildFileEntry(imagePath, tmpDir, multimodal);
+
+    const noteEntry = expectFileEntry(note);
+    expect(noteEntry.path).toBe("note.md");
+    expect(noteEntry.kind).toBe("markdown");
+    const imageEntry = expectFileEntry(image);
+    expect(imageEntry.path).toBe("diagram.png");
+    expect(imageEntry.kind).toBe("multimodal");
+    expect(imageEntry.modality).toBe("image");
+    expect(imageEntry.mimeType).toBe("image/png");
+    expect(imageEntry.contentText).toBe("Image file: diagram.png");
+  });
+
+  it("retries transient markdown reads while building file entries", async () => {
+    const tmpDir = getTmpDir();
+    const notePath = path.join(tmpDir, "note.md");
+    fsSync.writeFileSync(notePath, "hello", "utf-8");
+
+    const realOpen = fs.open;
+    let attempts = 0;
+    const openSpy = vi
+      .spyOn(fs, "open")
+      .mockImplementation(async (...args: Parameters<typeof realOpen>) => {
+        const [target, flags, mode] = args;
+        if (typeof target === "string" && path.resolve(target) === notePath && attempts++ === 0) {
+          const err = new Error(
+            "Unknown system error -11: Unknown system error -11, open",
+          ) as NodeJS.ErrnoException;
+          err.code = "UNKNOWN";
+          err.errno = -11;
+          throw err;
+        }
+        return await realOpen(target, flags, mode);
+      });
+
+    try {
+      const entry = expectFileEntry(await buildFileEntry(notePath, tmpDir));
+      expect(entry.path).toBe("note.md");
+      expect(entry.kind).toBe("markdown");
+      expect(attempts).toBe(2);
+    } finally {
+      openSpy.mockRestore();
+    }
+  });
+
+  it("builds multimodal chunks lazily and rejects changed files", async () => {
+    const tmpDir = getTmpDir();
+    const imagePath = path.join(tmpDir, "diagram.png");
+    fsSync.writeFileSync(imagePath, Buffer.from("png"));
+
+    const entry = expectFileEntry(await buildFileEntry(imagePath, tmpDir, multimodal));
+    const built = expectMultimodalIndexingChunk(await buildMultimodalChunkForIndexing(entry));
+    const parts = expectEmbeddingInput(built.chunk).parts ?? [];
+    expect(parts[0]).toEqual({ type: "text", text: "Image file: diagram.png" });
+    const inlinePart = parts[1];
+    if (inlinePart?.type !== "inline-data") {
+      throw new Error("Expected multimodal inline-data embedding part");
+    }
+    expect(inlinePart.mimeType).toBe("image/png");
+
+    fsSync.writeFileSync(imagePath, Buffer.alloc(entry.size + 32, 1));
+    await expect(buildMultimodalChunkForIndexing(entry)).resolves.toBeNull();
+  });
+
+  it("chunks mixed text and preserves surrogate pairs", () => {
+    const mixed = Array.from(
+      { length: 30 },
+      (_, index) => `Line ${index}: 这是中英文混合的测试内容 with English`,
+    ).join("\n");
+    const mixedChunks = chunkMarkdown(mixed, { tokens: 50, overlap: 0 });
+    expect(mixedChunks.length).toBeGreaterThan(1);
+    expect(mixedChunks.map((chunk) => chunk.text).join("\n")).toContain("Line 29");
+
+    const surrogateChar = "\u{20000}";
+    const surrogateChunks = chunkMarkdown(surrogateChar.repeat(120), {
+      tokens: 31,
+      overlap: 0,
+    });
+    for (const chunk of surrogateChunks) {
+      expect(chunk.text).not.toContain("\uFFFD");
+    }
+  });
+
+  it("remaps chunk lines using JSONL source line maps", () => {
+    const lineMap = [4, 6, 7, 10, 13];
+    const chunks = chunkMarkdown(
+      "User: Hello\nAssistant: Hi\nUser: Question\nAssistant: Answer\nUser: Thanks",
+      { tokens: 400, overlap: 0 },
+    );
+
+    remapChunkLines(chunks, lineMap);
+
+    expect(chunks[0].startLine).toBe(4);
+    expect(chunks[chunks.length - 1].endLine).toBe(13);
+  });
+});

packages/memory-host-sdk/src/host/internal.ts

@@ -0,0 +1,556 @@
+// Memory Host SDK module implements internal behavior.
+import crypto from "node:crypto";
+import fsSync from "node:fs";
+import fs from "node:fs/promises";
+import { homedir } from "node:os";
+import path from "node:path";
+import { CANONICAL_ROOT_MEMORY_FILENAME } from "./config-utils.js";
+import { estimateStructuredEmbeddingInputBytes } from "./embedding-input-limits.js";
+import { buildTextEmbeddingInput, type EmbeddingInput } from "./embedding-inputs.js";
+import {
+  isFileMissingError,
+  readRegularFile,
+  statRegularFile,
+  walkDirectory,
+  type WalkDirectoryEntry,
+} from "./fs-utils.js";
+import {
+  buildMemoryMultimodalLabel,
+  classifyMemoryMultimodalPath,
+  type MemoryMultimodalModality,
+  type MemoryMultimodalSettings,
+} from "./multimodal.js";
+import {
+  CHARS_PER_TOKEN_ESTIMATE,
+  detectMime,
+  estimateStringChars,
+  runTasksWithConcurrency,
+} from "./openclaw-runtime-io.js";
+import {
+  resolveCanonicalRootMemoryFile,
+  shouldSkipRootMemoryAuxiliaryPath,
+} from "./openclaw-runtime-memory.js";
+import { retryTransientMemoryRead } from "./read-retry.js";
+import { normalizeStringEntries, uniqueStrings } from "./string-utils.js";
+
+export { hashText } from "./hash.js";
+import { hashText } from "./hash.js";
+
+export type MemoryFileEntry = {
+  path: string;
+  absPath: string;
+  mtimeMs: number;
+  size: number;
+  hash: string;
+  dataHash?: string;
+  kind?: "markdown" | "multimodal";
+  contentText?: string;
+  modality?: MemoryMultimodalModality;
+  mimeType?: string;
+};
+
+export type MemoryChunk = {
+  startLine: number;
+  endLine: number;
+  text: string;
+  hash: string;
+  embeddingInput?: EmbeddingInput;
+};
+
+export type MultimodalMemoryChunk = {
+  chunk: MemoryChunk;
+  structuredInputBytes: number;
+};
+
+const DISABLED_MULTIMODAL_SETTINGS: MemoryMultimodalSettings = {
+  enabled: false,
+  modalities: [],
+  maxFileBytes: 0,
+};
+
+export function ensureDir(dir: string): string {
+  fsSync.mkdirSync(dir, { recursive: true });
+  return dir;
+}
+
+export function normalizeRelPath(value: string): string {
+  const trimmed = value.trim().replace(/^[./]+/, "");
+  return trimmed.replace(/\\/g, "/");
+}
+
+function expandHomePath(value: string): string {
+  if (value === "~") {
+    return homedir();
+  }
+  if (value.startsWith("~/") || value.startsWith("~\\")) {
+    return path.join(homedir(), value.slice(2));
+  }
+  return value;
+}
+
+export function normalizeExtraMemoryPaths(workspaceDir: string, extraPaths?: string[]): string[] {
+  if (!extraPaths?.length) {
+    return [];
+  }
+  const resolved = normalizeStringEntries(extraPaths)
+    .map((value) => expandHomePath(value))
+    .map((value) =>
+      path.isAbsolute(value) ? path.resolve(value) : path.resolve(workspaceDir, value),
+    );
+  return uniqueStrings(resolved);
+}
+
+export function isMemoryPath(relPath: string): boolean {
+  const normalized = normalizeRelPath(relPath);
+  if (!normalized) {
+    return false;
+  }
+  if (normalized === CANONICAL_ROOT_MEMORY_FILENAME || normalized.toLowerCase() === "dreams.md") {
+    return true;
+  }
+  return normalized.startsWith("memory/");
+}
+
+function isAllowedMemoryFilePath(filePath: string, multimodal?: MemoryMultimodalSettings): boolean {
+  if (filePath.endsWith(".md")) {
+    return true;
+  }
+  return (
+    classifyMemoryMultimodalPath(filePath, multimodal ?? DISABLED_MULTIMODAL_SETTINGS) !== null
+  );
+}
+
+function shouldDescendMemoryEntry(
+  entry: WalkDirectoryEntry,
+  shouldSkipPath?: (absPath: string) => boolean,
+): boolean {
+  if (shouldSkipPath?.(entry.path)) {
+    return false;
+  }
+  return entry.kind === "directory" && entry.name !== ".openclaw-repair";
+}
+
+async function collectMemoryFilesFromDir(
+  dir: string,
+  files: string[],
+  multimodal?: MemoryMultimodalSettings,
+  shouldSkipPath?: (absPath: string) => boolean,
+): Promise<void> {
+  const scan = await walkDirectory(dir, {
+    symlinks: "skip",
+    descend: (entry) => shouldDescendMemoryEntry(entry, shouldSkipPath),
+    include: (entry) =>
+      !shouldSkipPath?.(entry.path) &&
+      entry.kind === "file" &&
+      isAllowedMemoryFilePath(entry.path, multimodal),
+  });
+  files.push(...scan.entries.map((entry) => entry.path));
+}
+
+export async function listMemoryFiles(
+  workspaceDir: string,
+  extraPaths?: string[],
+  multimodal?: MemoryMultimodalSettings,
+): Promise<string[]> {
+  const result: string[] = [];
+  const memoryDir = path.join(workspaceDir, "memory");
+
+  const shouldSkipWorkspaceMemoryPath = (absPath: string): boolean =>
+    shouldSkipRootMemoryAuxiliaryPath({ workspaceDir, absPath });
+
+  const addMarkdownFile = async (absPath: string) => {
+    try {
+      const stat = await statRegularFile(absPath);
+      if (stat.missing) {
+        return;
+      }
+      if (!absPath.endsWith(".md")) {
+        return;
+      }
+      result.push(absPath);
+    } catch {}
+  };
+
+  const memoryFile = await resolveCanonicalRootMemoryFile(workspaceDir);
+  if (memoryFile) {
+    await addMarkdownFile(memoryFile);
+  }
+  try {
+    const dirStat = await fs.lstat(memoryDir);
+    if (!dirStat.isSymbolicLink() && dirStat.isDirectory()) {
+      await collectMemoryFilesFromDir(memoryDir, result, multimodal, shouldSkipWorkspaceMemoryPath);
+    }
+  } catch {}
+
+  const normalizedExtraPaths = normalizeExtraMemoryPaths(workspaceDir, extraPaths);
+  if (normalizedExtraPaths.length > 0) {
+    for (const inputPath of normalizedExtraPaths) {
+      if (shouldSkipWorkspaceMemoryPath(inputPath)) {
+        continue;
+      }
+      try {
+        const stat = await fs.lstat(inputPath);
+        if (stat.isSymbolicLink()) {
+          continue;
+        }
+        if (stat.isDirectory()) {
+          await collectMemoryFilesFromDir(
+            inputPath,
+            result,
+            multimodal,
+            shouldSkipWorkspaceMemoryPath,
+          );
+          continue;
+        }
+        if (stat.isFile() && isAllowedMemoryFilePath(inputPath, multimodal)) {
+          result.push(inputPath);
+        }
+      } catch {}
+    }
+  }
+  if (result.length <= 1) {
+    return result;
+  }
+  const seen = new Set<string>();
+  const deduped: string[] = [];
+  for (const entry of result) {
+    let key = entry;
+    try {
+      key = await fs.realpath(entry);
+    } catch {}
+    if (seen.has(key)) {
+      continue;
+    }
+    seen.add(key);
+    deduped.push(entry);
+  }
+  return deduped;
+}
+
+export async function buildFileEntry(
+  absPath: string,
+  workspaceDir: string,
+  multimodal?: MemoryMultimodalSettings,
+): Promise<MemoryFileEntry | null> {
+  const regularFile = await statRegularFile(absPath);
+  if (regularFile.missing) {
+    return null;
+  }
+  const stat = regularFile.stat;
+  const normalizedPath = path.relative(workspaceDir, absPath).replace(/\\/g, "/");
+  const multimodalSettings = multimodal ?? DISABLED_MULTIMODAL_SETTINGS;
+  const modality = classifyMemoryMultimodalPath(absPath, multimodalSettings);
+  if (modality) {
+    if (stat.size > multimodalSettings.maxFileBytes) {
+      return null;
+    }
+    let buffer: Buffer;
+    try {
+      buffer = (
+        await retryTransientMemoryRead(
+          () =>
+            readRegularFile({
+              filePath: absPath,
+              maxBytes: multimodalSettings.maxFileBytes,
+            }),
+          `read multimodal memory file ${absPath}`,
+        )
+      ).buffer;
+    } catch (err) {
+      if (isFileMissingError(err)) {
+        return null;
+      }
+      throw err;
+    }
+    const mimeType = await detectMime({ buffer: buffer.subarray(0, 512), filePath: absPath });
+    if (!mimeType || !mimeType.startsWith(`${modality}/`)) {
+      return null;
+    }
+    const contentText = buildMemoryMultimodalLabel(modality, normalizedPath);
+    const dataHash = crypto.createHash("sha256").update(buffer).digest("hex");
+    const chunkHash = hashText(
+      JSON.stringify({
+        path: normalizedPath,
+        contentText,
+        mimeType,
+        dataHash,
+      }),
+    );
+    return {
+      path: normalizedPath,
+      absPath,
+      mtimeMs: stat.mtimeMs,
+      size: stat.size,
+      hash: chunkHash,
+      dataHash,
+      kind: "multimodal",
+      contentText,
+      modality,
+      mimeType,
+    };
+  }
+  let content: string;
+  try {
+    content = (
+      await retryTransientMemoryRead(
+        () => readRegularFile({ filePath: absPath }),
+        `read memory index file ${absPath}`,
+      )
+    ).buffer.toString("utf-8");
+  } catch (err) {
+    if (isFileMissingError(err)) {
+      return null;
+    }
+    throw err;
+  }
+  const hash = hashText(content);
+  return {
+    path: normalizedPath,
+    absPath,
+    mtimeMs: stat.mtimeMs,
+    size: stat.size,
+    hash,
+    kind: "markdown",
+  };
+}
+
+async function loadMultimodalEmbeddingInput(
+  entry: Pick<
+    MemoryFileEntry,
+    "absPath" | "contentText" | "mimeType" | "kind" | "size" | "dataHash"
+  >,
+): Promise<EmbeddingInput | null> {
+  if (entry.kind !== "multimodal" || !entry.contentText || !entry.mimeType) {
+    return null;
+  }
+  const regularFile = await statRegularFile(entry.absPath);
+  if (regularFile.missing) {
+    return null;
+  }
+  const stat = regularFile.stat;
+  if (stat.size !== entry.size) {
+    return null;
+  }
+  let buffer: Buffer;
+  try {
+    buffer = (
+      await retryTransientMemoryRead(
+        () => readRegularFile({ filePath: entry.absPath, maxBytes: entry.size }),
+        `read multimodal indexing file ${entry.absPath}`,
+      )
+    ).buffer;
+  } catch (err) {
+    if (isFileMissingError(err)) {
+      return null;
+    }
+    throw err;
+  }
+  const dataHash = crypto.createHash("sha256").update(buffer).digest("hex");
+  if (entry.dataHash && entry.dataHash !== dataHash) {
+    return null;
+  }
+  return {
+    text: entry.contentText,
+    parts: [
+      { type: "text", text: entry.contentText },
+      {
+        type: "inline-data",
+        mimeType: entry.mimeType,
+        data: buffer.toString("base64"),
+      },
+    ],
+  };
+}
+
+export async function buildMultimodalChunkForIndexing(
+  entry: Pick<
+    MemoryFileEntry,
+    "absPath" | "contentText" | "mimeType" | "kind" | "hash" | "size" | "dataHash"
+  >,
+): Promise<MultimodalMemoryChunk | null> {
+  const embeddingInput = await loadMultimodalEmbeddingInput(entry);
+  if (!embeddingInput) {
+    return null;
+  }
+  return {
+    chunk: {
+      startLine: 1,
+      endLine: 1,
+      text: entry.contentText ?? embeddingInput.text,
+      hash: entry.hash,
+      embeddingInput,
+    },
+    structuredInputBytes: estimateStructuredEmbeddingInputBytes(embeddingInput),
+  };
+}
+
+export function chunkMarkdown(
+  content: string,
+  chunking: { tokens: number; overlap: number },
+): MemoryChunk[] {
+  const lines = content.split("\n");
+  if (lines.length === 0) {
+    return [];
+  }
+  const maxChars = Math.max(32, chunking.tokens * CHARS_PER_TOKEN_ESTIMATE);
+  const overlapChars = Math.max(0, chunking.overlap * CHARS_PER_TOKEN_ESTIMATE);
+  const chunks: MemoryChunk[] = [];
+
+  let current: Array<{ line: string; lineNo: number }> = [];
+  let currentChars = 0;
+
+  const flush = () => {
+    if (current.length === 0) {
+      return;
+    }
+    const firstEntry = current[0];
+    const lastEntry = current[current.length - 1];
+    if (!firstEntry || !lastEntry) {
+      return;
+    }
+    const text = current.map((entry) => entry.line).join("\n");
+    const startLine = firstEntry.lineNo;
+    const endLine = lastEntry.lineNo;
+    chunks.push({
+      startLine,
+      endLine,
+      text,
+      hash: hashText(text),
+      embeddingInput: buildTextEmbeddingInput(text),
+    });
+  };
+
+  const carryOverlap = () => {
+    if (overlapChars <= 0 || current.length === 0) {
+      current = [];
+      currentChars = 0;
+      return;
+    }
+    let acc = 0;
+    const kept: Array<{ line: string; lineNo: number }> = [];
+    for (let i = current.length - 1; i >= 0; i -= 1) {
+      const entry = current[i];
+      if (!entry) {
+        continue;
+      }
+      acc += estimateStringChars(entry.line) + 1;
+      kept.unshift(entry);
+      if (acc >= overlapChars) {
+        break;
+      }
+    }
+    current = kept;
+    currentChars = acc;
+  };
+
+  for (let i = 0; i < lines.length; i += 1) {
+    const line = lines[i] ?? "";
+    const lineNo = i + 1;
+    const segments: string[] = [];
+    if (line.length === 0) {
+      segments.push("");
+    } else {
+      // First pass: slice at maxChars (preserves original behaviour for Latin).
+      // Second pass: if a segment's *weighted* size still exceeds the budget
+      // (happens for CJK-heavy text where 1 char ≈ 1 token), re-split it at
+      // chunking.tokens so the chunk stays within the token budget.
+      for (let start = 0; start < line.length; start += maxChars) {
+        const coarse = line.slice(start, start + maxChars);
+        if (estimateStringChars(coarse) > maxChars) {
+          const fineStep = Math.max(1, chunking.tokens);
+          for (let j = 0; j < coarse.length; ) {
+            let end = Math.min(j + fineStep, coarse.length);
+            // Avoid splitting inside a UTF-16 surrogate pair (CJK Extension B+).
+            if (end < coarse.length) {
+              const code = coarse.charCodeAt(end - 1);
+              if (code >= 0xd800 && code <= 0xdbff) {
+                end += 1; // include the low surrogate
+              }
+            }
+            segments.push(coarse.slice(j, end));
+            j = end; // advance cursor to the adjusted boundary
+          }
+        } else {
+          segments.push(coarse);
+        }
+      }
+    }
+    for (const segment of segments) {
+      const lineSize = estimateStringChars(segment) + 1;
+      if (currentChars + lineSize > maxChars && current.length > 0) {
+        flush();
+        carryOverlap();
+      }
+      current.push({ line: segment, lineNo });
+      currentChars += lineSize;
+    }
+  }
+  flush();
+  return chunks;
+}
+
+/**
+ * Remap chunk startLine/endLine from content-relative positions to original
+ * source file positions using a lineMap.  Each entry in lineMap gives the
+ * 1-indexed source line for the corresponding 0-indexed content line.
+ *
+ * This is used for session JSONL files where buildSessionEntry() flattens
+ * messages into a plain-text string before chunking.  Without remapping the
+ * stored line numbers would reference positions in the flattened text rather
+ * than the original JSONL file.
+ */
+export function remapChunkLines(chunks: MemoryChunk[], lineMap: number[] | undefined): void {
+  if (!lineMap || lineMap.length === 0) {
+    return;
+  }
+  for (const chunk of chunks) {
+    // startLine/endLine are 1-indexed; lineMap is 0-indexed by content line
+    chunk.startLine = lineMap[chunk.startLine - 1] ?? chunk.startLine;
+    chunk.endLine = lineMap[chunk.endLine - 1] ?? chunk.endLine;
+  }
+}
+
+export function parseEmbedding(raw: string): number[] {
+  try {
+    const parsed = JSON.parse(raw) as number[];
+    return Array.isArray(parsed) ? parsed : [];
+  } catch {
+    return [];
+  }
+}
+
+export function cosineSimilarity(a: number[], b: number[]): number {
+  if (a.length === 0 || b.length === 0) {
+    return 0;
+  }
+  const len = Math.min(a.length, b.length);
+  let dot = 0;
+  let normA = 0;
+  let normB = 0;
+  for (let i = 0; i < len; i += 1) {
+    const av = a[i] ?? 0;
+    const bv = b[i] ?? 0;
+    dot += av * bv;
+    normA += av * av;
+    normB += bv * bv;
+  }
+  if (normA === 0 || normB === 0) {
+    return 0;
+  }
+  return dot / (Math.sqrt(normA) * Math.sqrt(normB));
+}
+
+export async function runWithConcurrency<T>(
+  tasks: Array<() => Promise<T>>,
+  limit: number,
+): Promise<T[]> {
+  const { results, firstError, hasError } = await runTasksWithConcurrency({
+    tasks,
+    limit,
+    errorMode: "stop",
+  });
+  if (hasError) {
+    throw firstError;
+  }
+  return results;
+}

packages/memory-host-sdk/src/host/memory-schema.test.ts

@@ -0,0 +1,340 @@
+// Memory schema tests cover canonical table creation and shipped-name migration.
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+import { DatabaseSync } from "node:sqlite";
+import { describe, expect, it } from "vitest";
+import { ensureMemoryIndexSchema } from "./memory-schema.js";
+
+describe("memory index schema", () => {
+  it("migrates shipped generic tables into canonical memory tables", () => {
+    const db = new DatabaseSync(":memory:");
+    try {
+      db.exec(`
+        CREATE TABLE meta (key TEXT PRIMARY KEY, value TEXT NOT NULL);
+        CREATE TABLE files (
+          path TEXT PRIMARY KEY,
+          source TEXT NOT NULL DEFAULT 'memory',
+          hash TEXT NOT NULL,
+          mtime INTEGER NOT NULL,
+          size INTEGER NOT NULL
+        );
+        CREATE TABLE chunks (
+          id TEXT PRIMARY KEY,
+          path TEXT NOT NULL,
+          source TEXT NOT NULL DEFAULT 'memory',
+          start_line INTEGER NOT NULL,
+          end_line INTEGER NOT NULL,
+          hash TEXT NOT NULL,
+          model TEXT NOT NULL,
+          text TEXT NOT NULL,
+          embedding TEXT NOT NULL,
+          updated_at INTEGER NOT NULL
+        );
+        CREATE TABLE embedding_cache (
+          provider TEXT NOT NULL,
+          model TEXT NOT NULL,
+          provider_key TEXT NOT NULL,
+          hash TEXT NOT NULL,
+          embedding TEXT NOT NULL,
+          dims INTEGER,
+          updated_at INTEGER NOT NULL,
+          PRIMARY KEY (provider, model, provider_key, hash)
+        );
+        CREATE VIRTUAL TABLE chunks_fts USING fts5(
+          text, id UNINDEXED, path UNINDEXED, source UNINDEXED, model UNINDEXED,
+          start_line UNINDEXED, end_line UNINDEXED
+        );
+        INSERT INTO meta VALUES ('memory_index_meta_v1', '{"vectorDims":3}');
+        INSERT INTO files VALUES ('MEMORY.md', 'memory', 'file-hash', 10, 20);
+        INSERT INTO chunks VALUES (
+          'chunk-1', 'MEMORY.md', 'memory', 1, 2, 'chunk-hash', 'embed-model',
+          'remember this', '[1,0,0]', 30
+        );
+        INSERT INTO embedding_cache VALUES (
+          'openai', 'embed-model', 'key', 'chunk-hash', '[1,0,0]', 3, 40
+        );
+        INSERT INTO chunks_fts VALUES (
+          'remember this', 'chunk-1', 'MEMORY.md', 'memory', 'embed-model', 1, 2
+        );
+      `);
+
+      const result = ensureMemoryIndexSchema({
+        db,
+        cacheEnabled: true,
+        ftsEnabled: true,
+      });
+
+      expect(result.ftsAvailable).toBe(true);
+      expect(db.prepare("SELECT * FROM memory_index_sources").all()).toEqual([
+        { path: "MEMORY.md", source: "memory", hash: "file-hash", mtime: 10, size: 20 },
+      ]);
+      expect(db.prepare("SELECT id, text FROM memory_index_chunks").all()).toEqual([
+        { id: "chunk-1", text: "remember this" },
+      ]);
+      expect(db.prepare("SELECT id, text FROM memory_index_chunks_fts").all()).toEqual([
+        { id: "chunk-1", text: "remember this" },
+      ]);
+      expect(db.prepare("SELECT provider, hash FROM memory_embedding_cache").all()).toEqual([
+        { provider: "openai", hash: "chunk-hash" },
+      ]);
+      expect(
+        db
+          .prepare(
+            "SELECT name FROM sqlite_master WHERE type = 'table' AND name IN ('meta', 'files', 'chunks', 'embedding_cache', 'chunks_fts')",
+          )
+          .all(),
+      ).toEqual([]);
+    } finally {
+      db.close();
+    }
+  });
+
+  it("does not import a legacy sidecar memory database during schema startup", () => {
+    const rootDir = fs.mkdtempSync(path.join(os.tmpdir(), "openclaw-memory-sidecar-"));
+    const legacyPath = path.join(rootDir, "memory", "main.sqlite");
+    const agentPath = path.join(rootDir, "agents", "main", "agent", "openclaw-agent.sqlite");
+    fs.mkdirSync(path.dirname(legacyPath), { recursive: true });
+    fs.mkdirSync(path.dirname(agentPath), { recursive: true });
+    const legacyDb = new DatabaseSync(legacyPath);
+    try {
+      legacyDb.exec(`
+        CREATE TABLE meta (key TEXT PRIMARY KEY, value TEXT NOT NULL);
+        CREATE TABLE files (
+          path TEXT PRIMARY KEY,
+          source TEXT NOT NULL DEFAULT 'memory',
+          hash TEXT NOT NULL,
+          mtime INTEGER NOT NULL,
+          size INTEGER NOT NULL
+        );
+        CREATE TABLE chunks (
+          id TEXT PRIMARY KEY,
+          path TEXT NOT NULL,
+          source TEXT NOT NULL DEFAULT 'memory',
+          start_line INTEGER NOT NULL,
+          end_line INTEGER NOT NULL,
+          hash TEXT NOT NULL,
+          model TEXT NOT NULL,
+          text TEXT NOT NULL,
+          embedding TEXT NOT NULL,
+          updated_at INTEGER NOT NULL
+        );
+        CREATE TABLE embedding_cache (
+          provider TEXT NOT NULL,
+          model TEXT NOT NULL,
+          provider_key TEXT NOT NULL,
+          hash TEXT NOT NULL,
+          embedding TEXT NOT NULL,
+          dims INTEGER,
+          updated_at INTEGER NOT NULL,
+          PRIMARY KEY (provider, model, provider_key, hash)
+        );
+        INSERT INTO meta VALUES ('memory_index_meta_v1', '{"vectorDims":3}');
+        INSERT INTO files VALUES ('MEMORY.md', 'memory', 'file-hash', 10, 20);
+        INSERT INTO chunks VALUES (
+          'chunk-1', 'MEMORY.md', 'memory', 1, 2, 'chunk-hash', 'embed-model',
+          'remember this', '[1,0,0]', 30
+        );
+        INSERT INTO embedding_cache VALUES (
+          'openai', 'embed-model', 'key', 'chunk-hash', '[1,0,0]', 3, 40
+        );
+      `);
+    } finally {
+      legacyDb.close();
+    }
+
+    const db = new DatabaseSync(agentPath);
+    try {
+      const result = ensureMemoryIndexSchema({
+        db,
+        cacheEnabled: true,
+        ftsEnabled: true,
+      });
+
+      expect(result.ftsAvailable).toBe(true);
+      expect(db.prepare("SELECT * FROM memory_index_sources").all()).toEqual([]);
+      expect(db.prepare("SELECT id, text FROM memory_index_chunks").all()).toEqual([]);
+      expect(db.prepare("SELECT id, text FROM memory_index_chunks_fts").all()).toEqual([]);
+      expect(db.prepare("SELECT provider, hash FROM memory_embedding_cache").all()).toEqual([]);
+      expect(fs.existsSync(legacyPath)).toBe(true);
+    } finally {
+      db.close();
+      fs.rmSync(rootDir, { recursive: true, force: true });
+    }
+  });
+
+  it("stores source records with the same path in separate sources", () => {
+    const db = new DatabaseSync(":memory:");
+    try {
+      ensureMemoryIndexSchema({
+        db,
+        cacheEnabled: false,
+        ftsEnabled: false,
+      });
+
+      db.prepare(
+        "INSERT INTO memory_index_sources (path, source, hash, mtime, size) VALUES (?, ?, ?, ?, ?)",
+      ).run("shared.md", "memory", "memory-hash", 10, 20);
+      db.prepare(
+        "INSERT INTO memory_index_sources (path, source, hash, mtime, size) VALUES (?, ?, ?, ?, ?)",
+      ).run("shared.md", "sessions", "session-hash", 30, 40);
+
+      expect(
+        db.prepare("SELECT path, source, hash FROM memory_index_sources ORDER BY source").all(),
+      ).toEqual([
+        { path: "shared.md", source: "memory", hash: "memory-hash" },
+        { path: "shared.md", source: "sessions", hash: "session-hash" },
+      ]);
+    } finally {
+      db.close();
+    }
+  });
+
+  it("honors shipped custom cache and FTS table names", () => {
+    const db = new DatabaseSync(":memory:");
+    try {
+      const result = ensureMemoryIndexSchema({
+        db,
+        embeddingCacheTable: "embedding_cache",
+        cacheEnabled: true,
+        ftsTable: "chunks_fts",
+        ftsEnabled: true,
+      });
+
+      expect(result.ftsAvailable).toBe(true);
+      expect(
+        db
+          .prepare(
+            "SELECT name FROM sqlite_master WHERE type = 'table' AND name IN ('embedding_cache', 'chunks_fts', 'memory_embedding_cache', 'memory_index_chunks_fts') ORDER BY name",
+          )
+          .all(),
+      ).toEqual([{ name: "chunks_fts" }, { name: "embedding_cache" }]);
+    } finally {
+      db.close();
+    }
+  });
+
+  it("upgrades canonical source tables keyed only by path", () => {
+    const db = new DatabaseSync(":memory:");
+    try {
+      db.exec(`
+        CREATE TABLE memory_index_sources (
+          path TEXT PRIMARY KEY,
+          source TEXT NOT NULL DEFAULT 'memory',
+          hash TEXT NOT NULL,
+          mtime INTEGER NOT NULL,
+          size INTEGER NOT NULL
+        );
+        INSERT INTO memory_index_sources VALUES ('shared.md', 'memory', 'memory-hash', 10, 20);
+      `);
+
+      ensureMemoryIndexSchema({
+        db,
+        cacheEnabled: false,
+        ftsEnabled: false,
+      });
+
+      db.prepare(
+        "INSERT INTO memory_index_sources (path, source, hash, mtime, size) VALUES (?, ?, ?, ?, ?)",
+      ).run("shared.md", "sessions", "session-hash", 30, 40);
+
+      expect(
+        db.prepare("SELECT path, source, hash FROM memory_index_sources ORDER BY source").all(),
+      ).toEqual([
+        { path: "shared.md", source: "memory", hash: "memory-hash" },
+        { path: "shared.md", source: "sessions", hash: "session-hash" },
+      ]);
+    } finally {
+      db.close();
+    }
+  });
+
+  it("leaves unrelated generic tables untouched", () => {
+    const db = new DatabaseSync(":memory:");
+    try {
+      db.exec(`
+        CREATE TABLE meta (key TEXT PRIMARY KEY, value TEXT NOT NULL, owner TEXT);
+        CREATE TABLE files (
+          path TEXT PRIMARY KEY,
+          source TEXT NOT NULL,
+          hash TEXT NOT NULL,
+          mtime INTEGER NOT NULL,
+          size INTEGER NOT NULL
+        );
+        CREATE TABLE chunks (
+          id TEXT PRIMARY KEY,
+          path TEXT NOT NULL,
+          source TEXT NOT NULL,
+          start_line INTEGER NOT NULL,
+          end_line INTEGER NOT NULL,
+          hash TEXT NOT NULL,
+          model TEXT NOT NULL,
+          text TEXT NOT NULL,
+          embedding TEXT NOT NULL,
+          updated_at INTEGER NOT NULL
+        );
+      `);
+
+      ensureMemoryIndexSchema({
+        db,
+        cacheEnabled: false,
+        ftsEnabled: false,
+      });
+
+      expect(
+        db
+          .prepare(
+            "SELECT name FROM sqlite_master WHERE type = 'table' AND name IN ('meta', 'files', 'chunks') ORDER BY name",
+          )
+          .all(),
+      ).toEqual([{ name: "chunks" }, { name: "files" }, { name: "meta" }]);
+    } finally {
+      db.close();
+    }
+  });
+
+  it("keeps legacy tables when canonical rows conflict", () => {
+    const db = new DatabaseSync(":memory:");
+    try {
+      db.exec(`
+        CREATE TABLE meta (key TEXT PRIMARY KEY, value TEXT NOT NULL);
+        CREATE TABLE files (
+          path TEXT PRIMARY KEY,
+          source TEXT NOT NULL DEFAULT 'memory',
+          hash TEXT NOT NULL,
+          mtime INTEGER NOT NULL,
+          size INTEGER NOT NULL
+        );
+        CREATE TABLE chunks (
+          id TEXT PRIMARY KEY,
+          path TEXT NOT NULL,
+          source TEXT NOT NULL DEFAULT 'memory',
+          start_line INTEGER NOT NULL,
+          end_line INTEGER NOT NULL,
+          hash TEXT NOT NULL,
+          model TEXT NOT NULL,
+          text TEXT NOT NULL,
+          embedding TEXT NOT NULL,
+          updated_at INTEGER NOT NULL
+        );
+        CREATE TABLE memory_index_meta (key TEXT PRIMARY KEY, value TEXT NOT NULL);
+        INSERT INTO meta VALUES ('memory_index_meta_v1', 'legacy');
+        INSERT INTO memory_index_meta VALUES ('memory_index_meta_v1', 'canonical');
+      `);
+
+      expect(() =>
+        ensureMemoryIndexSchema({
+          db,
+          cacheEnabled: false,
+          ftsEnabled: false,
+        }),
+      ).toThrow("legacy memory meta rows conflict");
+      expect(db.prepare("SELECT value FROM meta").get()).toEqual({ value: "legacy" });
+      expect(db.prepare("SELECT value FROM memory_index_meta").get()).toEqual({
+        value: "canonical",
+      });
+    } finally {
+      db.close();
+    }
+  });
+});

packages/memory-host-sdk/src/host/memory-schema.ts

@@ -0,0 +1,428 @@
+// Memory Host SDK module implements memory schema behavior.
+import type { DatabaseSync } from "node:sqlite";
+import { formatErrorMessage } from "./error-utils.js";
+
+// SQLite schema setup for builtin memory index, embedding cache, and FTS.
+
+export const MEMORY_INDEX_META_TABLE = "memory_index_meta";
+export const MEMORY_INDEX_SOURCES_TABLE = "memory_index_sources";
+export const MEMORY_INDEX_CHUNKS_TABLE = "memory_index_chunks";
+export const MEMORY_EMBEDDING_CACHE_TABLE = "memory_embedding_cache";
+export const MEMORY_INDEX_STATE_TABLE = "memory_index_state";
+export const MEMORY_INDEX_FTS_TABLE = "memory_index_chunks_fts";
+export const MEMORY_INDEX_VECTOR_TABLE = "memory_index_chunks_vec";
+
+const LEGACY_MEMORY_INDEX_TRIGGERS = [
+  "memory_files_revision_after_insert",
+  "memory_files_revision_after_update",
+  "memory_files_revision_after_delete",
+  "memory_chunks_revision_after_insert",
+  "memory_chunks_revision_after_update",
+  "memory_chunks_revision_after_delete",
+] as const;
+
+const MEMORY_INDEX_SOURCE_COLUMNS = ["path", "source", "hash", "mtime", "size"] as const;
+
+function tableColumns(db: DatabaseSync, tableName: string, schema = "main"): Set<string> {
+  const rows = db.prepare(`PRAGMA ${schema}.table_info(${tableName})`).all() as Array<{
+    name?: unknown;
+  }>;
+  return new Set(rows.flatMap((row) => (typeof row.name === "string" ? [row.name] : [])));
+}
+
+function tableHasExactColumns(
+  db: DatabaseSync,
+  tableName: string,
+  expected: readonly string[],
+  schema = "main",
+): boolean {
+  const columns = tableColumns(db, tableName, schema);
+  return columns.size === expected.length && expected.every((column) => columns.has(column));
+}
+
+function tablePrimaryKeyColumns(db: DatabaseSync, tableName: string): string[] {
+  const rows = db.prepare(`PRAGMA table_info(${tableName})`).all() as Array<{
+    name?: unknown;
+    pk?: unknown;
+  }>;
+  return rows
+    .flatMap((row) =>
+      typeof row.name === "string" && typeof row.pk === "number" && row.pk > 0
+        ? [{ name: row.name, pk: row.pk }]
+        : [],
+    )
+    .toSorted((left, right) => left.pk - right.pk)
+    .map((row) => row.name);
+}
+
+function tableHasPrimaryKey(
+  db: DatabaseSync,
+  tableName: string,
+  expectedColumns: readonly string[],
+): boolean {
+  const columns = tablePrimaryKeyColumns(db, tableName);
+  return (
+    columns.length === expectedColumns.length &&
+    columns.every((column, index) => column === expectedColumns[index])
+  );
+}
+
+function assertLegacyRowsCopied(db: DatabaseSync, query: string, tableName: string): void {
+  const row = db.prepare(query).get() as { missing?: unknown } | undefined;
+  if (Number(row?.missing ?? 0) > 0) {
+    throw new Error(`legacy memory ${tableName} rows conflict with canonical memory index rows`);
+  }
+}
+
+function migrateCanonicalMemoryIndexSourcesPrimaryKey(db: DatabaseSync): void {
+  if (
+    !tableHasExactColumns(db, MEMORY_INDEX_SOURCES_TABLE, MEMORY_INDEX_SOURCE_COLUMNS) ||
+    tableHasPrimaryKey(db, MEMORY_INDEX_SOURCES_TABLE, ["path", "source"])
+  ) {
+    return;
+  }
+  if (!tableHasPrimaryKey(db, MEMORY_INDEX_SOURCES_TABLE, ["path"])) {
+    return;
+  }
+
+  db.exec("SAVEPOINT migrate_memory_index_sources_primary_key");
+  try {
+    db.exec(`
+      DROP TRIGGER IF EXISTS memory_index_sources_revision_after_insert;
+      DROP TRIGGER IF EXISTS memory_index_sources_revision_after_update;
+      DROP TRIGGER IF EXISTS memory_index_sources_revision_after_delete;
+
+      ALTER TABLE ${MEMORY_INDEX_SOURCES_TABLE}
+        RENAME TO memory_index_sources_path_pk_migration;
+      CREATE TABLE ${MEMORY_INDEX_SOURCES_TABLE} (
+        path TEXT NOT NULL,
+        source TEXT NOT NULL DEFAULT 'memory',
+        hash TEXT NOT NULL,
+        mtime INTEGER NOT NULL,
+        size INTEGER NOT NULL,
+        PRIMARY KEY (path, source)
+      );
+      INSERT INTO ${MEMORY_INDEX_SOURCES_TABLE} (path, source, hash, mtime, size)
+      SELECT path, source, hash, mtime, size FROM memory_index_sources_path_pk_migration;
+      DROP TABLE memory_index_sources_path_pk_migration;
+      RELEASE migrate_memory_index_sources_primary_key;
+    `);
+  } catch (err) {
+    db.exec("ROLLBACK TO migrate_memory_index_sources_primary_key");
+    db.exec("RELEASE migrate_memory_index_sources_primary_key");
+    throw err;
+  }
+}
+
+function hasLegacyMemoryIndexTables(db: DatabaseSync, schema = "main"): boolean {
+  return (
+    tableHasExactColumns(db, "meta", ["key", "value"], schema) &&
+    tableHasExactColumns(db, "files", ["path", "source", "hash", "mtime", "size"], schema) &&
+    tableHasExactColumns(
+      db,
+      "chunks",
+      [
+        "id",
+        "path",
+        "source",
+        "start_line",
+        "end_line",
+        "hash",
+        "model",
+        "text",
+        "embedding",
+        "updated_at",
+      ],
+      schema,
+    )
+  );
+}
+
+function hasLegacyEmbeddingCacheTable(db: DatabaseSync, schema = "main"): boolean {
+  return tableHasExactColumns(
+    db,
+    "embedding_cache",
+    ["provider", "model", "provider_key", "hash", "embedding", "dims", "updated_at"],
+    schema,
+  );
+}
+
+function copyLegacyMemoryIndexRows(
+  db: DatabaseSync,
+  schema: string,
+  preservedEmbeddingCacheTable?: string,
+): void {
+  db.exec(`
+    INSERT OR IGNORE INTO main.${MEMORY_INDEX_META_TABLE} (key, value)
+    SELECT key, value FROM ${schema}.meta;
+
+    INSERT OR IGNORE INTO main.${MEMORY_INDEX_SOURCES_TABLE} (path, source, hash, mtime, size)
+    SELECT path, source, hash, mtime, size FROM ${schema}.files;
+
+    INSERT OR IGNORE INTO main.${MEMORY_INDEX_CHUNKS_TABLE} (
+      id, path, source, start_line, end_line, hash, model, text, embedding, updated_at
+    )
+    SELECT id, path, source, start_line, end_line, hash, model, text, embedding, updated_at
+    FROM ${schema}.chunks;
+  `);
+  assertLegacyRowsCopied(
+    db,
+    `SELECT COUNT(*) AS missing
+     FROM ${schema}.meta AS legacy
+     WHERE NOT EXISTS (
+       SELECT 1 FROM main.${MEMORY_INDEX_META_TABLE} AS canonical
+       WHERE canonical.key = legacy.key AND canonical.value IS legacy.value
+     )`,
+    "meta",
+  );
+  assertLegacyRowsCopied(
+    db,
+    `SELECT COUNT(*) AS missing
+     FROM ${schema}.files AS legacy
+     WHERE NOT EXISTS (
+       SELECT 1 FROM main.${MEMORY_INDEX_SOURCES_TABLE} AS canonical
+       WHERE canonical.path = legacy.path
+         AND canonical.source IS legacy.source
+         AND canonical.hash IS legacy.hash
+         AND canonical.mtime IS legacy.mtime
+         AND canonical.size IS legacy.size
+     )`,
+    "files",
+  );
+  assertLegacyRowsCopied(
+    db,
+    `SELECT COUNT(*) AS missing
+     FROM ${schema}.chunks AS legacy
+     WHERE NOT EXISTS (
+       SELECT 1 FROM main.${MEMORY_INDEX_CHUNKS_TABLE} AS canonical
+       WHERE canonical.id = legacy.id
+         AND canonical.path IS legacy.path
+         AND canonical.source IS legacy.source
+         AND canonical.start_line IS legacy.start_line
+         AND canonical.end_line IS legacy.end_line
+         AND canonical.hash IS legacy.hash
+         AND canonical.model IS legacy.model
+         AND canonical.text IS legacy.text
+         AND canonical.embedding IS legacy.embedding
+         AND canonical.updated_at IS legacy.updated_at
+     )`,
+    "chunks",
+  );
+  if (
+    preservedEmbeddingCacheTable !== "embedding_cache" &&
+    hasLegacyEmbeddingCacheTable(db, schema)
+  ) {
+    db.exec(`
+      CREATE TABLE IF NOT EXISTS main.${MEMORY_EMBEDDING_CACHE_TABLE} (
+        provider TEXT NOT NULL,
+        model TEXT NOT NULL,
+        provider_key TEXT NOT NULL,
+        hash TEXT NOT NULL,
+        embedding TEXT NOT NULL,
+        dims INTEGER,
+        updated_at INTEGER NOT NULL,
+        PRIMARY KEY (provider, model, provider_key, hash)
+      );
+      INSERT OR IGNORE INTO main.${MEMORY_EMBEDDING_CACHE_TABLE} (
+        provider, model, provider_key, hash, embedding, dims, updated_at
+      )
+      SELECT provider, model, provider_key, hash, embedding, dims, updated_at
+      FROM ${schema}.embedding_cache;
+    `);
+    assertLegacyRowsCopied(
+      db,
+      `SELECT COUNT(*) AS missing
+       FROM ${schema}.embedding_cache AS legacy
+       WHERE NOT EXISTS (
+         SELECT 1 FROM main.${MEMORY_EMBEDDING_CACHE_TABLE} AS canonical
+         WHERE canonical.provider = legacy.provider
+           AND canonical.model = legacy.model
+           AND canonical.provider_key = legacy.provider_key
+           AND canonical.hash = legacy.hash
+           AND canonical.embedding IS legacy.embedding
+           AND canonical.dims IS legacy.dims
+           AND canonical.updated_at IS legacy.updated_at
+       )`,
+      "embedding_cache",
+    );
+  }
+}
+
+function migrateLegacyMemoryIndexTables(
+  db: DatabaseSync,
+  preservedEmbeddingCacheTable?: string,
+): void {
+  if (!hasLegacyMemoryIndexTables(db)) {
+    return;
+  }
+
+  db.exec("SAVEPOINT migrate_legacy_memory_index_tables");
+  try {
+    copyLegacyMemoryIndexRows(db, "main", preservedEmbeddingCacheTable);
+    if (preservedEmbeddingCacheTable !== "embedding_cache" && hasLegacyEmbeddingCacheTable(db)) {
+      db.exec("DROP TABLE embedding_cache");
+    }
+    for (const trigger of LEGACY_MEMORY_INDEX_TRIGGERS) {
+      db.exec(`DROP TRIGGER IF EXISTS ${trigger}`);
+    }
+    db.exec(`
+      DROP TABLE IF EXISTS chunks_fts;
+      DROP TABLE chunks;
+      DROP TABLE files;
+      DROP TABLE meta;
+      RELEASE migrate_legacy_memory_index_tables;
+    `);
+  } catch (err) {
+    db.exec("ROLLBACK TO migrate_legacy_memory_index_tables");
+    db.exec("RELEASE migrate_legacy_memory_index_tables");
+    throw err;
+  }
+}
+
+/** Ensure canonical memory index tables and the optional FTS table exist. */
+export function ensureMemoryIndexSchema(params: {
+  db: DatabaseSync;
+  /** @deprecated Omit to use the canonical memory cache table. */
+  embeddingCacheTable?: string;
+  cacheEnabled: boolean;
+  /** @deprecated Omit to use the canonical memory FTS table. */
+  ftsTable?: string;
+  ftsEnabled: boolean;
+  ftsTokenizer?: "unicode61" | "trigram";
+}): { ftsAvailable: boolean; ftsError?: string } {
+  const embeddingCacheTable = params.embeddingCacheTable ?? MEMORY_EMBEDDING_CACHE_TABLE;
+  const ftsTable = params.ftsTable ?? MEMORY_INDEX_FTS_TABLE;
+  params.db.exec(`
+    CREATE TABLE IF NOT EXISTS ${MEMORY_INDEX_META_TABLE} (
+      key TEXT PRIMARY KEY,
+      value TEXT NOT NULL
+    );
+    CREATE TABLE IF NOT EXISTS ${MEMORY_INDEX_SOURCES_TABLE} (
+      path TEXT NOT NULL,
+      source TEXT NOT NULL DEFAULT 'memory',
+      hash TEXT NOT NULL,
+      mtime INTEGER NOT NULL,
+      size INTEGER NOT NULL,
+      PRIMARY KEY (path, source)
+    );
+    CREATE TABLE IF NOT EXISTS ${MEMORY_INDEX_CHUNKS_TABLE} (
+      id TEXT PRIMARY KEY,
+      path TEXT NOT NULL,
+      source TEXT NOT NULL DEFAULT 'memory',
+      start_line INTEGER NOT NULL,
+      end_line INTEGER NOT NULL,
+      hash TEXT NOT NULL,
+      model TEXT NOT NULL,
+      text TEXT NOT NULL,
+      embedding TEXT NOT NULL,
+      updated_at INTEGER NOT NULL
+    );
+    CREATE TABLE IF NOT EXISTS ${MEMORY_INDEX_STATE_TABLE} (
+      id INTEGER PRIMARY KEY CHECK (id = 1),
+      revision INTEGER NOT NULL
+    );
+    INSERT OR IGNORE INTO ${MEMORY_INDEX_STATE_TABLE} (id, revision) VALUES (1, 0);
+  `);
+  migrateCanonicalMemoryIndexSourcesPrimaryKey(params.db);
+  params.db.exec(`
+
+    CREATE TRIGGER IF NOT EXISTS memory_index_sources_revision_after_insert
+    AFTER INSERT ON ${MEMORY_INDEX_SOURCES_TABLE}
+    BEGIN
+      UPDATE ${MEMORY_INDEX_STATE_TABLE} SET revision = revision + 1 WHERE id = 1;
+    END;
+    CREATE TRIGGER IF NOT EXISTS memory_index_sources_revision_after_update
+    AFTER UPDATE ON ${MEMORY_INDEX_SOURCES_TABLE}
+    BEGIN
+      UPDATE ${MEMORY_INDEX_STATE_TABLE} SET revision = revision + 1 WHERE id = 1;
+    END;
+    CREATE TRIGGER IF NOT EXISTS memory_index_sources_revision_after_delete
+    AFTER DELETE ON ${MEMORY_INDEX_SOURCES_TABLE}
+    BEGIN
+      UPDATE ${MEMORY_INDEX_STATE_TABLE} SET revision = revision + 1 WHERE id = 1;
+    END;
+
+    CREATE TRIGGER IF NOT EXISTS memory_index_chunks_revision_after_insert
+    AFTER INSERT ON ${MEMORY_INDEX_CHUNKS_TABLE}
+    BEGIN
+      UPDATE ${MEMORY_INDEX_STATE_TABLE} SET revision = revision + 1 WHERE id = 1;
+    END;
+    CREATE TRIGGER IF NOT EXISTS memory_index_chunks_revision_after_update
+    AFTER UPDATE ON ${MEMORY_INDEX_CHUNKS_TABLE}
+    BEGIN
+      UPDATE ${MEMORY_INDEX_STATE_TABLE} SET revision = revision + 1 WHERE id = 1;
+    END;
+    CREATE TRIGGER IF NOT EXISTS memory_index_chunks_revision_after_delete
+    AFTER DELETE ON ${MEMORY_INDEX_CHUNKS_TABLE}
+    BEGIN
+      UPDATE ${MEMORY_INDEX_STATE_TABLE} SET revision = revision + 1 WHERE id = 1;
+    END;
+
+    CREATE INDEX IF NOT EXISTS idx_memory_index_sources_source
+      ON ${MEMORY_INDEX_SOURCES_TABLE}(source);
+    CREATE INDEX IF NOT EXISTS idx_memory_index_chunks_path_source
+      ON ${MEMORY_INDEX_CHUNKS_TABLE}(path, source);
+    CREATE INDEX IF NOT EXISTS idx_memory_index_chunks_path
+      ON ${MEMORY_INDEX_CHUNKS_TABLE}(path);
+    CREATE INDEX IF NOT EXISTS idx_memory_index_chunks_source
+      ON ${MEMORY_INDEX_CHUNKS_TABLE}(source);
+  `);
+  migrateLegacyMemoryIndexTables(params.db, params.embeddingCacheTable);
+  if (params.cacheEnabled) {
+    const updatedAtIndex =
+      embeddingCacheTable === MEMORY_EMBEDDING_CACHE_TABLE
+        ? "idx_memory_embedding_cache_updated_at"
+        : "idx_embedding_cache_updated_at";
+    params.db.exec(`
+      CREATE TABLE IF NOT EXISTS ${embeddingCacheTable} (
+        provider TEXT NOT NULL,
+        model TEXT NOT NULL,
+        provider_key TEXT NOT NULL,
+        hash TEXT NOT NULL,
+        embedding TEXT NOT NULL,
+        dims INTEGER,
+        updated_at INTEGER NOT NULL,
+        PRIMARY KEY (provider, model, provider_key, hash)
+      );
+      CREATE INDEX IF NOT EXISTS ${updatedAtIndex}
+        ON ${embeddingCacheTable}(updated_at);
+    `);
+  }
+
+  let ftsAvailable = false;
+  let ftsError: string | undefined;
+  if (params.ftsEnabled) {
+    try {
+      const tokenizer = params.ftsTokenizer ?? "unicode61";
+      const tokenizeClause = tokenizer === "trigram" ? `, tokenize='trigram case_sensitive 0'` : "";
+      params.db.exec(
+        `CREATE VIRTUAL TABLE IF NOT EXISTS ${ftsTable} USING fts5(\n` +
+          `  text,\n` +
+          `  id UNINDEXED,\n` +
+          `  path UNINDEXED,\n` +
+          `  source UNINDEXED,\n` +
+          `  model UNINDEXED,\n` +
+          `  start_line UNINDEXED,\n` +
+          `  end_line UNINDEXED\n` +
+          `${tokenizeClause});`,
+      );
+      // The shipped generic-table migration and a later FTS enablement both
+      // create an empty derived table beside already-canonical chunk rows.
+      params.db.exec(`
+        INSERT INTO ${ftsTable} (
+          text, id, path, source, model, start_line, end_line
+        )
+        SELECT text, id, path, source, model, start_line, end_line
+        FROM ${MEMORY_INDEX_CHUNKS_TABLE}
+        WHERE NOT EXISTS (SELECT 1 FROM ${ftsTable} LIMIT 1);
+      `);
+      ftsAvailable = true;
+    } catch (err) {
+      const message = formatErrorMessage(err);
+      ftsAvailable = false;
+      ftsError = message;
+    }
+  }
+
+  return { ftsAvailable, ...(ftsError ? { ftsError } : {}) };
+}

packages/memory-host-sdk/src/host/multimodal.ts

@@ -0,0 +1,117 @@
+// Memory Host SDK module implements multimodal behavior.
+import { normalizeLowercaseStringOrEmpty } from "./string-utils.js";
+
+// Multimodal memory settings and file classification helpers.
+
+const MEMORY_MULTIMODAL_SPECS = {
+  image: {
+    labelPrefix: "Image file",
+    extensions: [".jpg", ".jpeg", ".png", ".webp", ".gif", ".heic", ".heif"],
+  },
+  audio: {
+    labelPrefix: "Audio file",
+    extensions: [".mp3", ".wav", ".ogg", ".opus", ".m4a", ".aac", ".flac"],
+  },
+} as const;
+
+/** Supported multimodal memory modality. */
+export type MemoryMultimodalModality = keyof typeof MEMORY_MULTIMODAL_SPECS;
+/** All supported multimodal memory modalities in stable config order. */
+export const MEMORY_MULTIMODAL_MODALITIES = Object.keys(
+  MEMORY_MULTIMODAL_SPECS,
+) as MemoryMultimodalModality[];
+/** User selection for one modality or all modalities. */
+export type MemoryMultimodalSelection = MemoryMultimodalModality | "all";
+
+/** Normalized multimodal memory ingestion settings. */
+export type MemoryMultimodalSettings = {
+  enabled: boolean;
+  modalities: MemoryMultimodalModality[];
+  maxFileBytes: number;
+};
+
+/** Default max bytes for one multimodal memory file. */
+export const DEFAULT_MEMORY_MULTIMODAL_MAX_FILE_BYTES = 10 * 1024 * 1024;
+
+/** Normalize user modality selections to supported modalities. */
+export function normalizeMemoryMultimodalModalities(
+  raw: MemoryMultimodalSelection[] | undefined,
+): MemoryMultimodalModality[] {
+  if (raw === undefined || raw.includes("all")) {
+    return [...MEMORY_MULTIMODAL_MODALITIES];
+  }
+  const normalized = new Set<MemoryMultimodalModality>();
+  for (const value of raw) {
+    if (value === "image" || value === "audio") {
+      normalized.add(value);
+    }
+  }
+  return Array.from(normalized);
+}
+
+/** Normalize user multimodal settings, including disabled-state empty modality list. */
+export function normalizeMemoryMultimodalSettings(raw: {
+  enabled?: boolean;
+  modalities?: MemoryMultimodalSelection[];
+  maxFileBytes?: number;
+}): MemoryMultimodalSettings {
+  const enabled = raw.enabled === true;
+  const maxFileBytes =
+    typeof raw.maxFileBytes === "number" && Number.isFinite(raw.maxFileBytes)
+      ? Math.max(1, Math.floor(raw.maxFileBytes))
+      : DEFAULT_MEMORY_MULTIMODAL_MAX_FILE_BYTES;
+  return {
+    enabled,
+    modalities: enabled ? normalizeMemoryMultimodalModalities(raw.modalities) : [],
+    maxFileBytes,
+  };
+}
+
+/** Return true when multimodal memory ingestion has at least one enabled modality. */
+export function isMemoryMultimodalEnabled(settings: MemoryMultimodalSettings): boolean {
+  return settings.enabled && settings.modalities.length > 0;
+}
+
+/** Return accepted file extensions for a modality. */
+export function getMemoryMultimodalExtensions(
+  modality: MemoryMultimodalModality,
+): readonly string[] {
+  return MEMORY_MULTIMODAL_SPECS[modality].extensions;
+}
+
+/** Build the text label that accompanies embedded multimodal file content. */
+export function buildMemoryMultimodalLabel(
+  modality: MemoryMultimodalModality,
+  normalizedPath: string,
+): string {
+  return `${MEMORY_MULTIMODAL_SPECS[modality].labelPrefix}: ${normalizedPath}`;
+}
+
+/** Build a glob that matches an extension case-insensitively for QMD sources. */
+export function buildCaseInsensitiveExtensionGlob(extension: string): string {
+  const normalized = normalizeLowercaseStringOrEmpty(extension).replace(/^\./, "");
+  if (!normalized) {
+    return "*";
+  }
+  const parts = Array.from(normalized, (char) => `[${char.toLowerCase()}${char.toUpperCase()}]`);
+  return `*.${parts.join("")}`;
+}
+
+/** Classify a file path into a supported multimodal modality under current settings. */
+export function classifyMemoryMultimodalPath(
+  filePath: string,
+  settings: MemoryMultimodalSettings,
+): MemoryMultimodalModality | null {
+  if (!isMemoryMultimodalEnabled(settings)) {
+    return null;
+  }
+  const lower = normalizeLowercaseStringOrEmpty(filePath);
+  for (const modality of settings.modalities) {
+    for (const extension of getMemoryMultimodalExtensions(modality)) {
+      if (lower.endsWith(extension)) {
+        return modality;
+      }
+    }
+  }
+  return null;
+}

packages/memory-host-sdk/src/host/node-llama.ts

@@ -0,0 +1,52 @@
+// Minimal node-llama-cpp type facade used by the local embedding provider.
+
+/** Embedding vector returned by node-llama-cpp. */
+export type LlamaEmbedding = {
+  vector: Float32Array | number[];
+};
+
+/** Embedding context created from a loaded llama model. */
+export type LlamaEmbeddingContext = {
+  getEmbeddingFor: (text: string) => Promise<LlamaEmbedding>;
+  dispose?: () => Promise<void> | void;
+};
+
+/** Loaded llama model capable of creating embedding contexts. */
+export type LlamaModel = {
+  createEmbeddingContext: (options?: {
+    contextSize?: number | "auto";
+    createSignal?: AbortSignal;
+  }) => Promise<LlamaEmbeddingContext>;
+  dispose?: () => Promise<void> | void;
+};
+
+/** Options accepted by node-llama-cpp model file resolution. */
+export type ResolveModelFileOptions = {
+  directory?: string;
+  signal?: AbortSignal;
+};
+
+/** Root llama runtime object exposed by node-llama-cpp. */
+export type Llama = {
+  loadModel: (params: { modelPath: string; loadSignal?: AbortSignal }) => Promise<LlamaModel>;
+  dispose?: () => Promise<void> | void;
+};
+
+/** Imported node-llama-cpp module shape used by local embeddings. */
+export type NodeLlamaCppModule = {
+  LlamaLogLevel: {
+    error: number;
+  };
+  getLlama: (params: { logLevel: number }) => Promise<Llama>;
+  resolveModelFile: (
+    modelPath: string,
+    optionsOrDirectory?: string | ResolveModelFileOptions,
+  ) => Promise<string>;
+};
+
+const NODE_LLAMA_CPP_MODULE = "node-llama-cpp";
+
+/** Dynamically import node-llama-cpp so the optional dependency is loaded only when needed. */
+export async function importNodeLlamaCpp(moduleSpecifier = NODE_LLAMA_CPP_MODULE) {
+  return import(moduleSpecifier) as Promise<NodeLlamaCppModule>;
+}

packages/memory-host-sdk/src/host/openclaw-runtime-agent.ts

@@ -0,0 +1,23 @@
+// Agent-facing runtime facade for memory host packages.
+// Keep exports here limited to config/state helpers that memory plugins may reuse.
+export {
+  DEFAULT_AGENT_COMPACTION_RESERVE_TOKENS_FLOOR,
+  asToolParamsRecord,
+  jsonResult,
+  parseAgentSessionKey,
+  readNumberParam,
+  readStringParam,
+  resolveAgentContextLimits,
+  resolveAgentDir,
+  resolveAgentWorkspaceDir,
+  resolveCronStyleNow,
+  resolveDefaultAgentId,
+  resolveMemorySearchConfig,
+  resolveMemorySearchSyncConfig,
+  resolveSessionAgentId,
+} from "./openclaw-runtime.js";
+export type {
+  AnyAgentTool,
+  ResolvedMemorySearchConfig,
+  ResolvedMemorySearchSyncConfig,
+} from "./openclaw-runtime.js";

packages/memory-host-sdk/src/host/openclaw-runtime-auth.ts

@@ -0,0 +1,13 @@
+// Memory Host SDK module implements openclaw runtime auth behavior.
+import { requireApiKey } from "../../../../src/agents/model-auth-runtime-shared.js";
+import type { resolveApiKeyForProvider as ResolveApiKeyForProvider } from "../../../../src/agents/model-auth.js";
+
+// Lazy auth facade so memory host helpers avoid eager model-auth module loading.
+
+export { requireApiKey };
+
+/** Resolve a provider API key through the core model-auth runtime. */
+export const resolveApiKeyForProvider: typeof ResolveApiKeyForProvider = async (...args) => {
+  const auth = await import("../../../../src/agents/model-auth.js");
+  return auth.resolveApiKeyForProvider(...args);
+};

packages/memory-host-sdk/src/host/openclaw-runtime-cli.ts

@@ -0,0 +1,19 @@
+// Narrow CLI/runtime facade re-exported for memory host helpers.
+
+export {
+  colorize,
+  defaultRuntime,
+  formatDocsLink,
+  formatErrorMessage,
+  formatHelpExamples,
+  isRich,
+  isVerbose,
+  resolveCommandSecretRefsViaGateway,
+  setVerbose,
+  shortenHomeInString,
+  shortenHomePath,
+  theme,
+  withManager,
+  withProgress,
+  withProgressTotals,
+} from "./openclaw-runtime.js";

packages/memory-host-sdk/src/host/openclaw-runtime-config.ts

@@ -0,0 +1,24 @@
+// Config-facing runtime facade for memory host packages.
+// This keeps memory plugins off broader core config modules and their private helpers.
+export {
+  getRuntimeConfig,
+  hasConfiguredSecretInput,
+  loadConfig,
+  normalizeResolvedSecretInputString,
+  parseDurationMs,
+  parseNonNegativeByteSize,
+  resolveSessionTranscriptsDirForAgent,
+  resolveStateDir,
+} from "./openclaw-runtime.js";
+export type {
+  MemoryBackend,
+  MemoryCitationsMode,
+  MemoryQmdConfig,
+  MemoryQmdIndexPath,
+  MemoryQmdMcporterConfig,
+  MemoryQmdSearchMode,
+  MemorySearchConfig,
+  OpenClawConfig,
+  SecretInput,
+  SessionSendPolicyConfig,
+} from "./openclaw-runtime.js";

packages/memory-host-sdk/src/host/openclaw-runtime-io.ts

@@ -0,0 +1,43 @@
+// Narrow IO/runtime facade re-exported for memory host helpers.
+
+export {
+  CHARS_PER_TOKEN_ESTIMATE,
+  DEFAULT_SQLITE_WAL_AUTOCHECKPOINT_PAGES,
+  DEFAULT_SQLITE_WAL_CHECKPOINT_INTERVAL_MS,
+  DEFAULT_SQLITE_WAL_TRUNCATE_INTERVAL_MS,
+  applyWindowsSpawnProgramPolicy,
+  configureSqliteConnectionPragmas,
+  configureSqliteWalMaintenance,
+  root,
+  createSubsystemLogger,
+  detectMime,
+  estimateStringChars,
+  installProcessWarningFilter,
+  materializeWindowsSpawnProgram,
+  redactSensitiveText,
+  resolveGlobalSingleton,
+  resolveUserPath,
+  resolveWindowsExecutablePath,
+  resolveWindowsSpawnProgram,
+  resolveWindowsSpawnProgramCandidate,
+  runTasksWithConcurrency,
+  shortenHomeInString,
+  shortenHomePath,
+  shouldIgnoreWarning,
+  splitShellArgs,
+  truncateUtf16Safe,
+} from "./openclaw-runtime.js";
+
+export type {
+  ProcessWarning,
+  ResolveWindowsSpawnProgramCandidateParams,
+  ResolveWindowsSpawnProgramParams,
+  SqliteConnectionPragmaOptions,
+  SqliteWalMaintenance,
+  SqliteWalMaintenanceOptions,
+  WindowsSpawnCandidateResolution,
+  WindowsSpawnInvocation,
+  WindowsSpawnProgram,
+  WindowsSpawnProgramCandidate,
+  WindowsSpawnResolution,
+} from "./openclaw-runtime.js";

packages/memory-host-sdk/src/host/openclaw-runtime-memory.ts

@@ -0,0 +1,32 @@
+// Memory-facing runtime facade for plugin registration, embeddings, and prompt artifacts.
+// Re-export only stable host seams; plugin implementations should not import core internals.
+export {
+  buildActiveMemoryPromptSection,
+  emptyPluginConfigSchema,
+  getMemoryCapabilityRegistration,
+  getMemoryEmbeddingProvider,
+  listActiveMemoryPublicArtifacts,
+  listMemoryEmbeddingProviders,
+  listRegisteredMemoryEmbeddingProviderAdapters,
+  listRegisteredMemoryEmbeddingProviders,
+  resolveCanonicalRootMemoryFile,
+  shouldSkipRootMemoryAuxiliaryPath,
+} from "./openclaw-runtime.js";
+export type {
+  MemoryEmbeddingBatchChunk,
+  MemoryEmbeddingBatchOptions,
+  MemoryEmbeddingProvider,
+  MemoryEmbeddingProviderAdapter,
+  MemoryEmbeddingProviderCallOptions,
+  MemoryEmbeddingProviderCreateOptions,
+  MemoryEmbeddingProviderCreateResult,
+  MemoryEmbeddingProviderRuntime,
+  MemoryFlushPlan,
+  MemoryFlushPlanResolver,
+  MemoryPluginCapability,
+  MemoryPluginPublicArtifact,
+  MemoryPluginPublicArtifactsProvider,
+  MemoryPluginRuntime,
+  MemoryPromptSectionBuilder,
+  OpenClawPluginApi,
+} from "./openclaw-runtime.js";

packages/memory-host-sdk/src/host/openclaw-runtime-network.ts

@@ -0,0 +1,5 @@
+// Narrow network/runtime facade re-exported for memory remote HTTP helpers.
+
+export { fetchWithSsrFGuard } from "../../../../src/infra/net/fetch-guard.js";
+export { shouldUseEnvHttpProxyForUrl } from "../../../../src/infra/net/proxy-env.js";
+export { ssrfPolicyFromHttpBaseUrlAllowedHostname } from "../../../../src/infra/net/ssrf.js";

packages/memory-host-sdk/src/host/openclaw-runtime-session.ts

@@ -0,0 +1,62 @@
+// Narrow session/runtime facade re-exported for memory transcript helpers.
+import path from "node:path";
+
+export {
+  canonicalizeMainSessionAlias,
+  clearConfigCache,
+  clearRuntimeConfigSnapshot,
+  getRuntimeConfig,
+  HEARTBEAT_PROMPT,
+  HEARTBEAT_TOKEN,
+  SILENT_REPLY_TOKEN,
+  hasInterSessionUserProvenance,
+  isCompactionCheckpointTranscriptFileName,
+  isCronRunSessionKey,
+  isExecCompletionEvent,
+  isHeartbeatUserMessage,
+  isSessionArchiveArtifactName,
+  isSilentReplyPayloadText,
+  isUsageCountedSessionTranscriptFileName,
+  listSessionEntries,
+  onSessionTranscriptUpdate,
+  parseUsageCountedSessionIdFromFileName,
+  resolveSessionFilePath,
+  resolveStorePath,
+  resolveSessionAgentId,
+  resolveSessionTranscriptsDirForAgent,
+  stripInboundMetadata,
+  stripInternalRuntimeContext,
+  type SessionEntry,
+} from "./openclaw-runtime.js";
+
+/** Extracts the agent id from a canonical `agents/<id>/sessions` directory path. */
+export function extractAgentIdFromSessionsDir(sessionsDir: string): string | null {
+  const parts = path.normalize(path.resolve(sessionsDir)).split(path.sep).filter(Boolean);
+  const sessionsIndex = parts.length - 1;
+  if (
+    parts[sessionsIndex] !== "sessions" ||
+    sessionsIndex < 2 ||
+    parts[sessionsIndex - 2] !== "agents"
+  ) {
+    return null;
+  }
+  return parts[sessionsIndex - 1] || null;
+}
+
+/** Session-key prefix marking transcripts generated by memory dreaming runs. */
+export const DREAMING_NARRATIVE_RUN_PREFIX = "dreaming-narrative-";
+
+/** True when a session-store key belongs to a dreaming narrative run. */
+export function isDreamingNarrativeSessionStoreKey(sessionKey: string): boolean {
+  const trimmed = sessionKey.trim();
+  if (!trimmed) {
+    return false;
+  }
+  const firstSeparator = trimmed.indexOf(":");
+  if (firstSeparator < 0) {
+    return trimmed.startsWith(DREAMING_NARRATIVE_RUN_PREFIX);
+  }
+  const secondSeparator = trimmed.indexOf(":", firstSeparator + 1);
+  const sessionSegment = secondSeparator < 0 ? trimmed : trimmed.slice(secondSeparator + 1);
+  return sessionSegment.startsWith(DREAMING_NARRATIVE_RUN_PREFIX);
+}

packages/memory-host-sdk/src/host/openclaw-runtime.ts

@@ -0,0 +1,182 @@
+// Agent/runtime helpers.
+export { resolveCronStyleNow } from "../../../../src/agents/current-time.js";
+export {
+  resolveAgentContextLimits,
+  resolveAgentDir,
+  resolveAgentWorkspaceDir,
+  resolveDefaultAgentId,
+  resolveSessionAgentId,
+} from "../../../../src/agents/agent-scope.js";
+export { requireApiKey, resolveApiKeyForProvider } from "../../../../src/agents/model-auth.js";
+export { stripInternalRuntimeContext } from "../../../../src/agents/internal-runtime-context.js";
+export { DEFAULT_AGENT_COMPACTION_RESERVE_TOKENS_FLOOR } from "../../../../src/agents/agent-settings.js";
+export {
+  asToolParamsRecord,
+  jsonResult,
+  readNumberParam,
+  readStringParam,
+} from "../../../../src/agents/tools/common.js";
+export type { AnyAgentTool } from "../../../../src/agents/tools/common.js";
+export {
+  resolveMemorySearchConfig,
+  resolveMemorySearchSyncConfig,
+  type ResolvedMemorySearchConfig,
+  type ResolvedMemorySearchSyncConfig,
+} from "../../../../src/agents/memory-search.js";
+
+// Session and reply helpers.
+export { isHeartbeatUserMessage } from "../../../../src/auto-reply/heartbeat-filter.js";
+export { HEARTBEAT_PROMPT } from "../../../../src/auto-reply/heartbeat.js";
+export { stripInboundMetadata } from "../../../../src/auto-reply/reply/strip-inbound-meta.js";
+export {
+  HEARTBEAT_TOKEN,
+  SILENT_REPLY_TOKEN,
+  isSilentReplyPayloadText,
+} from "../../../../src/auto-reply/tokens.js";
+
+// CLI/runtime/config helpers.
+export { formatErrorMessage, withManager } from "../../../../src/cli/cli-utils.js";
+export { resolveCommandSecretRefsViaGateway } from "../../../../src/cli/command-secret-gateway.js";
+export { formatHelpExamples } from "../../../../src/cli/help-format.js";
+export { parseDurationMs } from "../../../../src/cli/parse-duration.js";
+export { withProgress, withProgressTotals } from "../../../../src/cli/progress.js";
+export { parseNonNegativeByteSize } from "../../../../src/config/byte-size.js";
+export {
+  clearConfigCache,
+  clearRuntimeConfigSnapshot,
+  getRuntimeConfig,
+  /** @deprecated Use getRuntimeConfig(), or pass the already loaded config through the call path. */
+  loadConfig,
+} from "../../../../src/config/config.js";
+export type { OpenClawConfig } from "../../../../src/config/config.js";
+export { resolveStateDir } from "../../../../src/config/paths.js";
+export {
+  isCompactionCheckpointTranscriptFileName,
+  isSessionArchiveArtifactName,
+  isUsageCountedSessionTranscriptFileName,
+  parseUsageCountedSessionIdFromFileName,
+} from "../../../../src/config/sessions/artifacts.js";
+export { canonicalizeMainSessionAlias } from "../../../../src/config/sessions/main-session.js";
+export { resolveSessionTranscriptsDirForAgent } from "../../../../src/config/sessions/paths.js";
+export {
+  listSessionEntries,
+  resolveSessionFilePath,
+  resolveStorePath,
+} from "../../../../src/plugin-sdk/session-store-runtime.js";
+export type { SessionEntry } from "../../../../src/config/sessions/types.js";
+export type { SessionSendPolicyConfig } from "../../../../src/config/types.base.js";
+export type {
+  MemoryBackend,
+  MemoryCitationsMode,
+  MemoryQmdConfig,
+  MemoryQmdIndexPath,
+  MemoryQmdMcporterConfig,
+  MemoryQmdSearchMode,
+} from "../../../../src/config/types.memory.js";
+export {
+  hasConfiguredSecretInput,
+  normalizeResolvedSecretInputString,
+} from "../../../../src/config/types.secrets.js";
+export type { SecretInput } from "../../../../src/config/types.secrets.js";
+export type { MemorySearchConfig } from "../../../../src/config/types.tools.js";
+export { isVerbose, setVerbose } from "../../../../src/globals.js";
+
+// IO, network, and logging helpers.
+export { isExecCompletionEvent } from "../../../../src/infra/heartbeat-events-filter.js";
+export { root } from "../../../../src/infra/fs-safe.js";
+export { fetchWithSsrFGuard } from "../../../../src/infra/net/fetch-guard.js";
+export { shouldUseEnvHttpProxyForUrl } from "../../../../src/infra/net/proxy-env.js";
+export { ssrfPolicyFromHttpBaseUrlAllowedHostname } from "../../../../src/infra/net/ssrf.js";
+export {
+  DEFAULT_SQLITE_WAL_AUTOCHECKPOINT_PAGES,
+  DEFAULT_SQLITE_WAL_CHECKPOINT_INTERVAL_MS,
+  DEFAULT_SQLITE_WAL_TRUNCATE_INTERVAL_MS,
+  configureSqliteConnectionPragmas,
+  configureSqliteWalMaintenance,
+} from "../../../../src/infra/sqlite-wal.js";
+export type {
+  SqliteConnectionPragmaOptions,
+  SqliteWalMaintenance,
+  SqliteWalMaintenanceOptions,
+} from "../../../../src/infra/sqlite-wal.js";
+export {
+  installProcessWarningFilter,
+  shouldIgnoreWarning,
+} from "../../../../src/infra/warning-filter.js";
+export type { ProcessWarning } from "../../../../src/infra/warning-filter.js";
+export { redactSensitiveText } from "../../../../src/logging/redact.js";
+export { createSubsystemLogger } from "../../../../src/logging/subsystem.js";
+export { detectMime } from "@openclaw/media-core/mime";
+
+// Memory plugin helpers.
+export {
+  resolveCanonicalRootMemoryFile,
+  shouldSkipRootMemoryAuxiliaryPath,
+} from "../../../../src/memory/root-memory-files.js";
+export {
+  getMemoryEmbeddingProvider,
+  listMemoryEmbeddingProviders,
+  listRegisteredMemoryEmbeddingProviderAdapters,
+  listRegisteredMemoryEmbeddingProviders,
+} from "../../../../src/plugins/memory-embedding-provider-runtime.js";
+export type {
+  MemoryEmbeddingBatchChunk,
+  MemoryEmbeddingBatchOptions,
+  MemoryEmbeddingProvider,
+  MemoryEmbeddingProviderAdapter,
+  MemoryEmbeddingProviderCallOptions,
+  MemoryEmbeddingProviderCreateOptions,
+  MemoryEmbeddingProviderCreateResult,
+  MemoryEmbeddingProviderRuntime,
+} from "../../../../src/plugins/memory-embedding-providers.js";
+export { emptyPluginConfigSchema } from "../../../../src/plugins/config-schema.js";
+export {
+  buildMemoryPromptSection as buildActiveMemoryPromptSection,
+  getMemoryCapabilityRegistration,
+  listActiveMemoryPublicArtifacts,
+} from "../../../../src/plugins/memory-state.js";
+export type {
+  MemoryFlushPlan,
+  MemoryFlushPlanResolver,
+  MemoryPluginCapability,
+  MemoryPluginPublicArtifact,
+  MemoryPluginPublicArtifactsProvider,
+  MemoryPluginRuntime,
+  MemoryPromptSectionBuilder,
+} from "../../../../src/plugins/memory-state.js";
+export type { OpenClawPluginApi } from "../../../../src/plugins/types.js";
+
+// Shared session/text utilities.
+export { defaultRuntime } from "../../../../src/runtime.js";
+export { parseAgentSessionKey } from "../../../../src/routing/session-key.js";
+export { hasInterSessionUserProvenance } from "../../../../src/sessions/input-provenance.js";
+export { isCronRunSessionKey } from "../../../../src/sessions/session-key-utils.js";
+export { onSessionTranscriptUpdate } from "../../../../src/sessions/transcript-events.js";
+export { formatDocsLink } from "../../../terminal-core/src/links.js";
+export { colorize, isRich, theme } from "../../../terminal-core/src/theme.js";
+export { CHARS_PER_TOKEN_ESTIMATE, estimateStringChars } from "../../../../src/utils/cjk-chars.js";
+export { runTasksWithConcurrency } from "../../../../src/utils/run-with-concurrency.js";
+export { splitShellArgs } from "../../../../src/utils/shell-argv.js";
+export {
+  resolveUserPath,
+  shortenHomeInString,
+  shortenHomePath,
+  truncateUtf16Safe,
+} from "../../../../src/utils.js";
+export {
+  applyWindowsSpawnProgramPolicy,
+  materializeWindowsSpawnProgram,
+  resolveWindowsExecutablePath,
+  resolveWindowsSpawnProgram,
+  resolveWindowsSpawnProgramCandidate,
+} from "../../../../src/plugin-sdk/windows-spawn.js";
+export type {
+  ResolveWindowsSpawnProgramCandidateParams,
+  ResolveWindowsSpawnProgramParams,
+  WindowsSpawnCandidateResolution,
+  WindowsSpawnInvocation,
+  WindowsSpawnProgram,
+  WindowsSpawnProgramCandidate,
+  WindowsSpawnResolution,
+} from "../../../../src/plugin-sdk/windows-spawn.js";
+export { resolveGlobalSingleton } from "../../../../src/shared/global-singleton.js";

packages/memory-host-sdk/src/host/post-json.test.ts

@@ -0,0 +1,242 @@
+// Memory Host SDK tests cover post json behavior.
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import { postJson } from "./post-json.js";
+import { withRemoteHttpResponse } from "./remote-http.js";
+
+vi.mock("./remote-http.js", () => ({
+  withRemoteHttpResponse: vi.fn(),
+}));
+
+const remoteHttpMock = vi.mocked(withRemoteHttpResponse);
+
+function jsonResponse(payload: unknown, status = 200): Response {
+  return new Response(JSON.stringify(payload), { status });
+}
+
+function textResponse(body: string, status: number): Response {
+  return new Response(body, { status });
+}
+
+function streamingTextResponse(params: {
+  body: string;
+  status: number;
+  headers?: HeadersInit;
+  onCancel: () => void;
+}): Response {
+  const encoded = new TextEncoder().encode(params.body);
+  const stream = new ReadableStream<Uint8Array>({
+    start(controller) {
+      controller.enqueue(encoded);
+    },
+    cancel() {
+      params.onCancel();
+    },
+  });
+  return new Response(stream, { status: params.status, headers: params.headers });
+}
+
+function stallingSuccessResponse(onCancel: () => void): Response {
+  const reader = {
+    read: () => new Promise<ReadableStreamReadResult<Uint8Array>>(() => {}),
+    cancel: async () => {
+      onCancel();
+    },
+    releaseLock: () => undefined,
+  } as ReadableStreamDefaultReader<Uint8Array>;
+
+  return {
+    body: { getReader: () => reader },
+    headers: new Headers(),
+    ok: true,
+    status: 200,
+  } as Response;
+}
+
+describe("postJson", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+  });
+
+  it("parses JSON payload on successful response", async () => {
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(jsonResponse({ data: [{ embedding: [1, 2] }] }));
+    });
+
+    const result = await postJson({
+      url: "https://memory.example/v1/post",
+      headers: { Authorization: "Bearer test" },
+      body: { input: ["x"] },
+      errorPrefix: "post failed",
+      parse: (payload) => payload,
+    });
+
+    expect(result).toEqual({ data: [{ embedding: [1, 2] }] });
+  });
+
+  it("forwards abort signals to the remote HTTP request", async () => {
+    const controller = new AbortController();
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      expect(params.signal).toBe(controller.signal);
+      return await params.onResponse(jsonResponse({ ok: true }));
+    });
+
+    await postJson({
+      url: "https://memory.example/v1/post",
+      headers: {},
+      body: {},
+      signal: controller.signal,
+      errorPrefix: "post failed",
+      parse: (payload) => payload,
+    });
+  });
+
+  it("applies abort signals while reading successful response bodies", async () => {
+    let canceled = false;
+    const controller = new AbortController();
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(
+        stallingSuccessResponse(() => {
+          canceled = true;
+        }),
+      );
+    });
+
+    const read = postJson({
+      url: "https://memory.example/v1/post",
+      headers: {},
+      body: {},
+      signal: controller.signal,
+      errorPrefix: "post failed",
+      parse: () => ({}),
+    });
+
+    await new Promise((resolve) => {
+      setTimeout(resolve, 0);
+    });
+    controller.abort(new Error("body aborted"));
+
+    await expect(read).rejects.toThrow("body aborted");
+    expect(canceled).toBe(true);
+  });
+
+  it("attaches status to thrown error when requested", async () => {
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(textResponse("bad gateway", 502));
+    });
+
+    let error: unknown;
+    try {
+      await postJson({
+        url: "https://memory.example/v1/post",
+        headers: {},
+        body: {},
+        errorPrefix: "post failed",
+        attachStatus: true,
+        parse: () => ({}),
+      });
+    } catch (caught) {
+      error = caught;
+    }
+
+    expect(error).toBeInstanceOf(Error);
+    expect((error as Error).message).toBe("post failed: 502 bad gateway");
+    expect((error as { status?: unknown }).status).toBe(502);
+  });
+
+  it("bounds non-ok response bodies before formatting the error", async () => {
+    let canceled = false;
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(
+        streamingTextResponse({
+          body: "x".repeat(12_000),
+          status: 502,
+          onCancel: () => {
+            canceled = true;
+          },
+        }),
+      );
+    });
+
+    await expect(
+      postJson({
+        url: "https://memory.example/v1/post",
+        headers: {},
+        body: {},
+        errorPrefix: "post failed",
+        parse: () => ({}),
+      }),
+    ).rejects.toThrow(`post failed: 502 ${"x".repeat(1_000)}... [truncated]`);
+    expect(canceled).toBe(true);
+  });
+
+  it("wraps malformed success JSON with the request error prefix", async () => {
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(textResponse("{ nope", 200));
+    });
+
+    await expect(
+      postJson({
+        url: "https://memory.example/v1/post",
+        headers: {},
+        body: {},
+        errorPrefix: "post failed",
+        parse: () => ({}),
+      }),
+    ).rejects.toThrow("post failed: malformed JSON response");
+  });
+
+  it("rejects successful JSON responses with oversized content-length", async () => {
+    let canceled = false;
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(
+        streamingTextResponse({
+          body: "{}",
+          status: 200,
+          headers: { "content-length": "32" },
+          onCancel: () => {
+            canceled = true;
+          },
+        }),
+      );
+    });
+
+    await expect(
+      postJson({
+        url: "https://memory.example/v1/post",
+        headers: {},
+        body: {},
+        errorPrefix: "post failed",
+        maxResponseBytes: 8,
+        parse: () => ({}),
+      }),
+    ).rejects.toThrow("post failed: response body too large: 32 bytes (limit: 8 bytes)");
+    expect(canceled).toBe(true);
+  });
+
+  it("cancels successful JSON responses that exceed the streaming byte cap", async () => {
+    let canceled = false;
+    remoteHttpMock.mockImplementationOnce(async (params) => {
+      return await params.onResponse(
+        streamingTextResponse({
+          body: `{"data":"${"x".repeat(32)}"}`,
+          status: 200,
+          onCancel: () => {
+            canceled = true;
+          },
+        }),
+      );
+    });
+
+    await expect(
+      postJson({
+        url: "https://memory.example/v1/post",
+        headers: {},
+        body: {},
+        errorPrefix: "post failed",
+        maxResponseBytes: 16,
+        parse: () => ({}),
+      }),
+    ).rejects.toThrow("post failed: response body too large");
+    expect(canceled).toBe(true);
+  });
+});

packages/memory-host-sdk/src/host/post-json.ts

@@ -0,0 +1,50 @@
+// Memory Host SDK module implements post json behavior.
+import { withRemoteHttpResponse } from "./remote-http.js";
+import { readResponseJsonWithLimit, readResponseTextSnippet } from "./response-snippet.js";
+import type { SsrFPolicy } from "./ssrf-policy.js";
+
+// Shared JSON POST helper for guarded remote memory provider calls.
+
+/** POST JSON, parse bounded response JSON, and attach status metadata when requested. */
+export async function postJson<T>(params: {
+  url: string;
+  headers: Record<string, string>;
+  ssrfPolicy?: SsrFPolicy;
+  fetchImpl?: typeof fetch;
+  signal?: AbortSignal;
+  body: unknown;
+  errorPrefix: string;
+  attachStatus?: boolean;
+  maxResponseBytes?: number;
+  parse: (payload: unknown) => T | Promise<T>;
+}): Promise<T> {
+  return await withRemoteHttpResponse({
+    url: params.url,
+    ssrfPolicy: params.ssrfPolicy,
+    fetchImpl: params.fetchImpl,
+    signal: params.signal,
+    init: {
+      method: "POST",
+      headers: params.headers,
+      body: JSON.stringify(params.body),
+    },
+    onResponse: async (res) => {
+      if (!res.ok) {
+        const text = await readResponseTextSnippet(res, { signal: params.signal });
+        const err = new Error(`${params.errorPrefix}: ${res.status} ${text}`) as Error & {
+          status?: number;
+        };
+        if (params.attachStatus) {
+          err.status = res.status;
+        }
+        throw err;
+      }
+      const payload = await readResponseJsonWithLimit(res, {
+        errorPrefix: params.errorPrefix,
+        maxBytes: params.maxResponseBytes,
+        signal: params.signal,
+      });
+      return await params.parse(payload);
+    },
+  });
+}

packages/memory-host-sdk/src/host/qmd-process.real.test.ts

@@ -0,0 +1,143 @@
+// Memory Host SDK real-process tests cover QMD process-tree cleanup.
+import { spawnSync } from "node:child_process";
+import fs from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { describe, expect, it } from "vitest";
+import { runCliCommand } from "./qmd-process.js";
+
+type ProcessTreePids = {
+  parent: number;
+  grandchild: number;
+};
+
+function isProcessRunning(pid: number): boolean {
+  try {
+    process.kill(pid, 0);
+    return true;
+  } catch (error) {
+    if (error instanceof Error && "code" in error && error.code === "ESRCH") {
+      return false;
+    }
+    throw error;
+  }
+}
+
+async function waitUntil(params: {
+  condition: () => boolean | Promise<boolean>;
+  description: string;
+  timeoutMs?: number;
+}): Promise<void> {
+  const deadline = Date.now() + (params.timeoutMs ?? 5_000);
+  while (!(await params.condition())) {
+    if (Date.now() >= deadline) {
+      throw new Error(`timed out waiting for ${params.description}`);
+    }
+    await new Promise<void>((resolve) => {
+      setTimeout(resolve, 20);
+    });
+  }
+}
+
+async function readProcessTreePids(pidFile: string): Promise<ProcessTreePids> {
+  let pids: ProcessTreePids | undefined;
+  await waitUntil({
+    description: "the process-tree PID file",
+    condition: async () => {
+      try {
+        pids = JSON.parse(await fs.readFile(pidFile, "utf8")) as ProcessTreePids;
+        return Number.isInteger(pids.parent) && Number.isInteger(pids.grandchild);
+      } catch (error) {
+        if (error instanceof Error && "code" in error && error.code === "ENOENT") {
+          return false;
+        }
+        throw error;
+      }
+    },
+  });
+  if (!pids) {
+    throw new Error("process-tree PID file was not populated");
+  }
+  return pids;
+}
+
+function killProcessTree(parentPid: number): void {
+  if (process.platform === "win32") {
+    const systemRoot = process.env.SystemRoot ?? process.env.WINDIR ?? "C:\\Windows";
+    spawnSync(
+      path.win32.join(systemRoot, "System32", "taskkill.exe"),
+      ["/PID", String(parentPid), "/T", "/F"],
+      { stdio: "ignore", windowsHide: true },
+    );
+    return;
+  }
+  // The production abort path already force-kills the group. Cleanup after a
+  // failed assertion starts gracefully so it cannot kill a reused group id.
+  process.kill(-parentPid, "SIGTERM");
+}
+
+describe("runCliCommand real process lifecycle", () => {
+  it("kills the command and its descendant when the caller aborts", async () => {
+    const tempDir = await fs.mkdtemp(path.join(os.tmpdir(), "openclaw-qmd-abort-"));
+    const pidFile = path.join(tempDir, "pids.json");
+    const controller = new AbortController();
+    const abortError = new Error("memory_search timed out after 15s");
+    let pending: ReturnType<typeof runCliCommand> | undefined;
+    let pids: ProcessTreePids | undefined;
+
+    const childScript = `
+        const { spawn } = require("node:child_process");
+        const { renameSync, writeFileSync } = require("node:fs");
+        const grandchild = spawn(process.execPath, ["-e", "setInterval(() => {}, 1000)"], {
+          stdio: "ignore",
+        });
+        const pidFile = process.argv[1];
+        const temporaryPidFile = pidFile + ".tmp";
+        writeFileSync(temporaryPidFile, JSON.stringify({
+          parent: process.pid,
+          grandchild: grandchild.pid,
+        }));
+        renameSync(temporaryPidFile, pidFile);
+        setInterval(() => {}, 1000);
+      `;
+
+    try {
+      pending = runCliCommand({
+        commandSummary: "real qmd process-tree fixture",
+        spawnInvocation: { command: process.execPath, argv: ["-e", childScript, pidFile] },
+        env: process.env,
+        cwd: tempDir,
+        timeoutMs: 60_000,
+        maxOutputChars: 10_000,
+        signal: controller.signal,
+      });
+
+      pids = await readProcessTreePids(pidFile);
+      expect(isProcessRunning(pids.parent)).toBe(true);
+      expect(isProcessRunning(pids.grandchild)).toBe(true);
+
+      controller.abort(abortError);
+      await expect(pending).rejects.toBe(abortError);
+      await waitUntil({
+        description: "the detached process tree to exit",
+        condition: () =>
+          pids !== undefined &&
+          !isProcessRunning(pids.parent) &&
+          !isProcessRunning(pids.grandchild),
+      });
+    } finally {
+      if (!controller.signal.aborted) {
+        controller.abort(abortError);
+      }
+      await pending?.catch(() => undefined);
+      if (pids?.parent && isProcessRunning(pids.parent)) {
+        try {
+          killProcessTree(pids.parent);
+        } catch {
+          // Best-effort cleanup when an assertion failed after the child exited.
+        }
+      }
+      await fs.rm(tempDir, { recursive: true, force: true });
+    }
+  }, 15_000);
+});

packages/memory-host-sdk/src/host/qmd-process.test.ts

@@ -0,0 +1,527 @@
+// Memory Host SDK tests cover qmd process behavior.
+import { EventEmitter } from "node:events";
+import fs from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import {
+  afterAll,
+  afterEach,
+  beforeAll,
+  beforeEach,
+  describe,
+  expect,
+  it,
+  vi,
+  type MockInstance,
+} from "vitest";
+import { MAX_SAFE_TIMEOUT_DELAY_MS } from "../../../gateway-client/src/timeouts.js";
+
+const spawnMock = vi.hoisted(() => vi.fn());
+const spawnSyncMock = vi.hoisted(() => vi.fn());
+
+vi.mock("node:child_process", async () => {
+  const actual = await vi.importActual<typeof import("node:child_process")>("node:child_process");
+  return {
+    ...actual,
+    spawn: spawnMock,
+    spawnSync: spawnSyncMock,
+  };
+});
+
+import {
+  checkQmdBinaryAvailability,
+  resolveCliSpawnInvocation,
+  resolveQmdBinaryUnavailableReason,
+  runCliCommand,
+  type QmdBinaryAvailability,
+} from "./qmd-process.js";
+
+function createMockChild(params: { pid?: number } = {}) {
+  const child = new EventEmitter() as EventEmitter & {
+    pid?: number;
+    stdout: EventEmitter;
+    stderr: EventEmitter;
+    kill: ReturnType<typeof vi.fn>;
+    closeWith: (code?: number | null, signal?: NodeJS.Signals | null) => void;
+  };
+  child.pid = params.pid;
+  child.stdout = new EventEmitter();
+  child.stderr = new EventEmitter();
+  child.kill = vi.fn();
+  child.closeWith = (code: number | null = 0, signal: NodeJS.Signals | null = null) => {
+    child.emit("close", code, signal);
+  };
+  return child;
+}
+
+let fixtureRoot = "";
+let tempDir = "";
+let platformSpy: MockInstance<() => NodeJS.Platform> | null = null;
+let fixtureId = 0;
+const originalPath = process.env.PATH;
+const originalPathExt = process.env.PATHEXT;
+const originalSystemRoot = process.env.SystemRoot;
+const originalWindir = process.env.WINDIR;
+const taskkillPath = path.win32.join("C:\\Windows", "System32", "taskkill.exe");
+
+function restoreEnvValue(key: string, value: string | undefined): void {
+  if (value === undefined) {
+    delete process.env[key];
+    return;
+  }
+  process.env[key] = value;
+}
+
+beforeAll(async () => {
+  fixtureRoot = await fs.mkdtemp(path.join(os.tmpdir(), "openclaw-qmd-win-spawn-"));
+  platformSpy = vi.spyOn(process, "platform", "get").mockReturnValue("win32");
+});
+
+afterAll(async () => {
+  platformSpy?.mockRestore();
+  platformSpy = null;
+  if (fixtureRoot) {
+    await fs.rm(fixtureRoot, { recursive: true, force: true });
+  }
+});
+
+beforeEach(async () => {
+  tempDir = path.join(fixtureRoot, `case-${fixtureId++}`);
+  await fs.mkdir(tempDir, { recursive: true });
+  process.env.SystemRoot = "C:\\Windows";
+  delete process.env.WINDIR;
+});
+
+afterEach(() => {
+  vi.useRealTimers();
+  process.env.PATH = originalPath;
+  process.env.PATHEXT = originalPathExt;
+  restoreEnvValue("SystemRoot", originalSystemRoot);
+  restoreEnvValue("WINDIR", originalWindir);
+  platformSpy?.mockReturnValue("win32");
+  spawnMock.mockReset();
+  spawnSyncMock.mockReset();
+  spawnSyncMock.mockReturnValue({ status: 0 });
+  tempDir = "";
+});
+
+describe("resolveCliSpawnInvocation", () => {
+  it("unwraps npm cmd shims to a direct node entrypoint", async () => {
+    const binDir = path.join(tempDir, "node_modules", ".bin");
+    const packageDir = path.join(tempDir, "node_modules", "qmd");
+    const scriptPath = path.join(packageDir, "dist", "cli.js");
+    await fs.mkdir(path.dirname(scriptPath), { recursive: true });
+    await fs.mkdir(binDir, { recursive: true });
+    await fs.writeFile(path.join(binDir, "qmd.cmd"), "@echo off\r\n", "utf8");
+    await fs.writeFile(
+      path.join(packageDir, "package.json"),
+      JSON.stringify({ name: "qmd", version: "0.0.0", bin: { qmd: "dist/cli.js" } }),
+      "utf8",
+    );
+    await fs.writeFile(scriptPath, "module.exports = {};\n", "utf8");
+
+    process.env.PATH = `${binDir};${originalPath ?? ""}`;
+    process.env.PATHEXT = ".CMD;.EXE";
+
+    const invocation = resolveCliSpawnInvocation({
+      command: "qmd",
+      args: ["query", "hello"],
+      env: process.env,
+      packageName: "qmd",
+    });
+
+    expect(invocation.command).toBe(process.execPath);
+    expect(invocation.argv).toEqual([scriptPath, "query", "hello"]);
+    expect(invocation.shell).not.toBe(true);
+    expect(invocation.windowsHide).toBe(true);
+  });
+
+  it("fails closed when a Windows cmd shim cannot be resolved without shell execution", async () => {
+    const binDir = path.join(tempDir, "bad-bin");
+    await fs.mkdir(binDir, { recursive: true });
+    await fs.writeFile(path.join(binDir, "qmd.cmd"), "@echo off\r\nREM no entrypoint\r\n", "utf8");
+
+    process.env.PATH = `${binDir};${originalPath ?? ""}`;
+    process.env.PATHEXT = ".CMD;.EXE";
+
+    expect(() =>
+      resolveCliSpawnInvocation({
+        command: "qmd",
+        args: ["query", "hello"],
+        env: process.env,
+        packageName: "qmd",
+      }),
+    ).toThrow(/without shell execution/);
+  });
+
+  it("keeps bare commands bare when no Windows wrapper exists on PATH", () => {
+    process.env.PATH = originalPath ?? "";
+    process.env.PATHEXT = ".CMD;.EXE";
+
+    const invocation = resolveCliSpawnInvocation({
+      command: "qmd",
+      args: ["query", "hello"],
+      env: process.env,
+      packageName: "qmd",
+    });
+
+    expect(invocation.command).toBe("qmd");
+    expect(invocation.argv).toEqual(["query", "hello"]);
+    expect(invocation.shell).not.toBe(true);
+  });
+});
+
+describe("checkQmdBinaryAvailability", () => {
+  it("keeps legacy unavailable probe results source-compatible", () => {
+    const legacyUnavailable: QmdBinaryAvailability = {
+      available: false,
+      error: "spawn qmd ENOENT",
+    };
+
+    expect(resolveQmdBinaryUnavailableReason(legacyUnavailable)).toBe("binary");
+  });
+
+  it("returns available when the qmd process spawns successfully", async () => {
+    const child = createMockChild({ pid: 12344 });
+    spawnMock.mockImplementationOnce(() => {
+      queueMicrotask(() => child.emit("spawn"));
+      return child;
+    });
+
+    await expect(
+      checkQmdBinaryAvailability({ command: "qmd", env: process.env, cwd: tempDir }),
+    ).resolves.toEqual({ available: true });
+    expect(spawnSyncMock).toHaveBeenCalledWith(taskkillPath, ["/PID", String(child.pid), "/T"], {
+      stdio: "ignore",
+      windowsHide: true,
+    });
+    expect(child.kill).not.toHaveBeenCalled();
+  });
+
+  it("force-kills Windows availability probes when graceful taskkill fails", async () => {
+    const child = createMockChild({ pid: 12345 });
+    spawnMock.mockImplementationOnce(() => {
+      queueMicrotask(() => child.emit("spawn"));
+      return child;
+    });
+    spawnSyncMock.mockReset();
+    spawnSyncMock.mockReturnValueOnce({ status: 1 }).mockReturnValueOnce({ status: 0 });
+
+    await expect(
+      checkQmdBinaryAvailability({ command: "qmd", env: process.env, cwd: tempDir }),
+    ).resolves.toEqual({ available: true });
+
+    expect(spawnSyncMock).toHaveBeenNthCalledWith(1, taskkillPath, ["/PID", "12345", "/T"], {
+      stdio: "ignore",
+      windowsHide: true,
+    });
+    expect(spawnSyncMock).toHaveBeenNthCalledWith(2, taskkillPath, ["/PID", "12345", "/T", "/F"], {
+      stdio: "ignore",
+      windowsHide: true,
+    });
+    expect(child.kill).not.toHaveBeenCalled();
+  });
+
+  it("returns unavailable when the qmd process cannot be spawned", async () => {
+    const child = createMockChild();
+    const err = Object.assign(new Error("spawn qmd ENOENT"), { code: "ENOENT" });
+    spawnMock.mockImplementationOnce(() => {
+      queueMicrotask(() => child.emit("error", err));
+      return child;
+    });
+
+    await expect(
+      checkQmdBinaryAvailability({ command: "qmd", env: process.env, cwd: tempDir }),
+    ).resolves.toEqual({ available: false, reason: "binary", error: "spawn qmd ENOENT" });
+  });
+
+  it("returns an explicit workspace error when cwd is missing", async () => {
+    const missingDir = path.join(tempDir, "missing-workspace");
+
+    await expect(
+      checkQmdBinaryAvailability({ command: "qmd", env: process.env, cwd: missingDir }),
+    ).resolves.toEqual({
+      available: false,
+      reason: "workspace-cwd",
+      error: `workspace directory missing: ${missingDir}`,
+    });
+    expect(spawnMock).not.toHaveBeenCalled();
+  });
+
+  it("does not treat close-before-spawn as a successful availability probe", async () => {
+    const child = createMockChild();
+    const err = Object.assign(new Error("spawn qmd ENOENT"), { code: "ENOENT" });
+    spawnMock.mockImplementationOnce(() => {
+      queueMicrotask(() => child.emit("close"));
+      queueMicrotask(() => child.emit("error", err));
+      return child;
+    });
+
+    await expect(
+      checkQmdBinaryAvailability({ command: "qmd", env: process.env, cwd: tempDir }),
+    ).resolves.toEqual({ available: false, reason: "binary", error: "spawn qmd ENOENT" });
+  });
+
+  it("caps oversized availability probe timeouts before scheduling", async () => {
+    vi.useFakeTimers();
+    const child = createMockChild();
+    const timeoutSpy = vi.spyOn(globalThis, "setTimeout");
+    spawnMock.mockReturnValueOnce(child);
+
+    void checkQmdBinaryAvailability({
+      command: "qmd",
+      env: process.env,
+      cwd: tempDir,
+      timeoutMs: Number.MAX_SAFE_INTEGER,
+    });
+
+    expect(timeoutSpy).toHaveBeenCalledWith(expect.any(Function), MAX_SAFE_TIMEOUT_DELAY_MS);
+  });
+  it("kills timed-out availability probes by process group on POSIX", async () => {
+    platformSpy?.mockReturnValue("linux");
+    const killProcess = vi.spyOn(process, "kill").mockImplementation(() => true);
+    const child = createMockChild({ pid: 4321 });
+    spawnMock.mockReturnValueOnce(child);
+
+    try {
+      await expect(
+        checkQmdBinaryAvailability({
+          command: "qmd",
+          env: process.env,
+          cwd: tempDir,
+          timeoutMs: 1,
+        }),
+      ).resolves.toEqual({
+        available: false,
+        reason: "binary",
+        error: "spawn qmd timed out after 1ms",
+      });
+
+      expect(spawnMock.mock.calls[0]?.[2]).toMatchObject({ detached: true });
+      expect(killProcess).toHaveBeenCalledWith(-4321, "SIGKILL");
+      expect(child.kill).not.toHaveBeenCalledWith("SIGKILL");
+    } finally {
+      killProcess.mockRestore();
+    }
+  });
+});
+
+describe("runCliCommand", () => {
+  it("keeps stdout and stderr on non-zero exits", async () => {
+    const child = createMockChild();
+    spawnMock.mockImplementationOnce(() => {
+      queueMicrotask(() => {
+        child.stdout.emit("data", '[{"docid":"abc","score":0.93}]');
+        child.stderr.emit("data", "ggml-metal-device.m:612");
+        child.closeWith(134);
+      });
+      return child;
+    });
+
+    try {
+      await runCliCommand({
+        commandSummary: "qmd query test",
+        spawnInvocation: { command: "qmd", argv: ["query", "test", "--json"] },
+        env: process.env,
+        cwd: tempDir,
+        maxOutputChars: 10_000,
+      });
+      throw new Error("expected runCliCommand to reject");
+    } catch (err) {
+      expect(err).toBeInstanceOf(Error);
+      if (!(err instanceof Error)) {
+        throw err;
+      }
+      expect(err.name).toBe("CliCommandError");
+      expect(err).toMatchObject({
+        code: 134,
+        signal: null,
+        stdout: '[{"docid":"abc","score":0.93}]',
+        stderr: "ggml-metal-device.m:612",
+      });
+      expect(err.message).toContain("qmd query test failed (code 134)");
+    }
+  });
+
+  it("records signal-only command failures", async () => {
+    const child = createMockChild();
+    spawnMock.mockImplementationOnce(() => {
+      queueMicrotask(() => {
+        child.stdout.emit("data", "[]");
+        child.closeWith(null, "SIGABRT");
+      });
+      return child;
+    });
+
+    await expect(
+      runCliCommand({
+        commandSummary: "qmd query test",
+        spawnInvocation: { command: "qmd", argv: ["query", "test", "--json"] },
+        env: process.env,
+        cwd: tempDir,
+        maxOutputChars: 10_000,
+      }),
+    ).rejects.toMatchObject({
+      code: null,
+      signal: "SIGABRT",
+      stdout: "[]",
+    });
+  });
+
+  it("does not expose truncated output as a recoverable command failure", async () => {
+    const child = createMockChild();
+    spawnMock.mockImplementationOnce(() => {
+      queueMicrotask(() => {
+        child.stdout.emit("data", "too much output");
+        child.closeWith(1);
+      });
+      return child;
+    });
+
+    await expect(
+      runCliCommand({
+        commandSummary: "qmd query test",
+        spawnInvocation: { command: "qmd", argv: ["query", "test", "--json"] },
+        env: process.env,
+        cwd: tempDir,
+        maxOutputChars: 4,
+      }),
+    ).rejects.toThrow(/produced too much output/);
+  });
+
+  it("counts surrogate pairs as one character when capping failed command output", async () => {
+    const child = createMockChild();
+    spawnMock.mockImplementationOnce(() => {
+      queueMicrotask(() => {
+        child.stderr.emit("data", "a🙂");
+        child.closeWith(1);
+      });
+      return child;
+    });
+
+    await expect(
+      runCliCommand({
+        commandSummary: "qmd query test",
+        spawnInvocation: { command: "qmd", argv: ["query", "test", "--json"] },
+        env: process.env,
+        cwd: tempDir,
+        maxOutputChars: 2,
+      }),
+    ).rejects.toThrow(/🙂/);
+  });
+
+  it("caps oversized command timeouts before scheduling", async () => {
+    vi.useFakeTimers();
+    const child = createMockChild();
+    const timeoutSpy = vi.spyOn(globalThis, "setTimeout");
+    spawnMock.mockReturnValueOnce(child);
+
+    void runCliCommand({
+      commandSummary: "qmd query test",
+      spawnInvocation: { command: "qmd", argv: ["query", "test", "--json"] },
+      env: process.env,
+      cwd: tempDir,
+      maxOutputChars: 10_000,
+      timeoutMs: Number.MAX_SAFE_INTEGER,
+    }).catch(() => undefined);
+
+    expect(timeoutSpy).toHaveBeenCalledWith(expect.any(Function), MAX_SAFE_TIMEOUT_DELAY_MS);
+  });
+
+  it("kills aborted cli command process groups on POSIX and rejects with the abort reason", async () => {
+    platformSpy?.mockReturnValue("linux");
+    const killProcess = vi.spyOn(process, "kill").mockImplementation(() => true);
+    const child = createMockChild({ pid: 7654 });
+    spawnMock.mockReturnValueOnce(child);
+    const controller = new AbortController();
+
+    try {
+      const pending = runCliCommand({
+        commandSummary: "qmd query slow",
+        spawnInvocation: { command: "qmd", argv: ["query", "slow", "--json"] },
+        env: process.env,
+        cwd: tempDir,
+        maxOutputChars: 10_000,
+        timeoutMs: 60_000,
+        signal: controller.signal,
+      });
+
+      controller.abort(new Error("memory_search timed out after 15s"));
+
+      await expect(pending).rejects.toThrow("memory_search timed out after 15s");
+      expect(spawnMock.mock.calls[0]?.[2]).toMatchObject({ detached: true });
+      expect(killProcess).toHaveBeenCalledWith(-7654, "SIGKILL");
+      expect(child.kill).not.toHaveBeenCalledWith("SIGKILL");
+    } finally {
+      killProcess.mockRestore();
+    }
+  });
+
+  it("rejects immediately without spawning when the signal is already aborted", async () => {
+    const controller = new AbortController();
+    controller.abort(new Error("memory_search timed out after 15s"));
+
+    await expect(
+      runCliCommand({
+        commandSummary: "qmd query test",
+        spawnInvocation: { command: "qmd", argv: ["query", "test", "--json"] },
+        env: process.env,
+        cwd: tempDir,
+        maxOutputChars: 10_000,
+        signal: controller.signal,
+      }),
+    ).rejects.toThrow("memory_search timed out after 15s");
+    expect(spawnMock).not.toHaveBeenCalled();
+  });
+
+  it("kills timed-out cli command process groups on POSIX", async () => {
+    platformSpy?.mockReturnValue("linux");
+    const killProcess = vi.spyOn(process, "kill").mockImplementation(() => true);
+    const child = createMockChild({ pid: 8765 });
+    spawnMock.mockReturnValueOnce(child);
+
+    try {
+      const pending = runCliCommand({
+        commandSummary: "qmd query test",
+        spawnInvocation: { command: "qmd", argv: ["query", "test", "--json"] },
+        env: process.env,
+        cwd: tempDir,
+        maxOutputChars: 10_000,
+        timeoutMs: 1,
+      });
+      const timeoutAssertion = expect(pending).rejects.toThrow(
+        "qmd query test timed out after 1ms",
+      );
+
+      await timeoutAssertion;
+
+      expect(spawnMock.mock.calls[0]?.[2]).toMatchObject({ detached: true });
+      expect(killProcess).toHaveBeenCalledWith(-8765, "SIGKILL");
+      expect(child.kill).not.toHaveBeenCalledWith("SIGKILL");
+    } finally {
+      killProcess.mockRestore();
+    }
+  });
+
+  it("force-kills timed-out Windows cli commands with taskkill", async () => {
+    const child = createMockChild({ pid: 12346 });
+    spawnMock.mockReturnValueOnce(child);
+
+    const pending = runCliCommand({
+      commandSummary: "qmd query test",
+      spawnInvocation: { command: "qmd", argv: ["query", "test", "--json"] },
+      env: process.env,
+      cwd: tempDir,
+      maxOutputChars: 10_000,
+      timeoutMs: 1,
+    });
+
+    await expect(pending).rejects.toThrow("qmd query test timed out after 1ms");
+
+    expect(spawnSyncMock).toHaveBeenCalledWith(taskkillPath, ["/PID", "12346", "/T", "/F"], {
+      stdio: "ignore",
+      windowsHide: true,
+    });
+    expect(child.kill).not.toHaveBeenCalledWith("SIGKILL");
+  });
+});

packages/memory-host-sdk/src/host/qmd-process.ts

@@ -0,0 +1,425 @@
+// Memory Host SDK module implements qmd process behavior.
+import { spawn, spawnSync } from "node:child_process";
+import { statSync } from "node:fs";
+import path from "node:path";
+import { resolveSafeTimeoutDelayMs } from "../../../gateway-client/src/timeouts.js";
+import { materializeWindowsSpawnProgram, resolveWindowsSpawnProgram } from "./windows-spawn.js";
+
+export type CliSpawnInvocation = {
+  command: string;
+  argv: string[];
+  shell?: boolean;
+  windowsHide?: boolean;
+};
+
+type QmdChildProcess = {
+  pid?: number;
+  kill: (signal?: NodeJS.Signals) => boolean;
+};
+
+const DEFAULT_WINDOWS_SYSTEM_ROOT = "C:\\Windows";
+
+export type QmdBinaryUnavailableReason = "binary" | "workspace-cwd";
+
+export type QmdBinaryUnavailable = {
+  available: false;
+  /**
+   * Optional for source compatibility with older plugin SDK callers that
+   * returned only `{ available: false, error }`.
+   */
+  reason?: QmdBinaryUnavailableReason;
+  error: string;
+};
+
+export type QmdBinaryAvailability = { available: true } | QmdBinaryUnavailable;
+
+export function resolveQmdBinaryUnavailableReason(
+  result: QmdBinaryUnavailable,
+): QmdBinaryUnavailableReason {
+  return result.reason ?? "binary";
+}
+
+export function resolveCliSpawnInvocation(params: {
+  command: string;
+  args: string[];
+  env: NodeJS.ProcessEnv;
+  packageName: string;
+}): CliSpawnInvocation {
+  const program = resolveWindowsSpawnProgram({
+    command: params.command,
+    platform: process.platform,
+    env: params.env,
+    execPath: process.execPath,
+    packageName: params.packageName,
+    allowShellFallback: false,
+  });
+  return materializeWindowsSpawnProgram(program, params.args);
+}
+
+export async function checkQmdBinaryAvailability(params: {
+  command: string;
+  env: NodeJS.ProcessEnv;
+  cwd?: string;
+  timeoutMs?: number;
+}): Promise<QmdBinaryAvailability> {
+  let spawnInvocation: CliSpawnInvocation;
+  try {
+    spawnInvocation = resolveCliSpawnInvocation({
+      command: params.command,
+      args: [],
+      env: params.env,
+      packageName: "qmd",
+    });
+  } catch (err) {
+    return { available: false, reason: "binary", error: formatQmdAvailabilityError(err) };
+  }
+
+  const cwd = params.cwd ?? process.cwd();
+  const cwdError = validateQmdProbeCwd(cwd);
+  if (cwdError) {
+    return cwdError;
+  }
+
+  return await new Promise((resolve) => {
+    let settled = false;
+    let didSpawn = false;
+    const finish = (result: QmdBinaryAvailability) => {
+      if (settled) {
+        return;
+      }
+      settled = true;
+      if (timer) {
+        clearTimeout(timer);
+      }
+      resolve(result);
+    };
+
+    const child = spawn(spawnInvocation.command, spawnInvocation.argv, {
+      env: params.env,
+      cwd,
+      shell: spawnInvocation.shell,
+      windowsHide: spawnInvocation.windowsHide,
+      stdio: "ignore",
+      detached: shouldUseQmdProcessGroup(),
+    });
+    const timeoutMs = resolveSafeTimeoutDelayMs(params.timeoutMs ?? 2_000, { minMs: 0 });
+    const timer = setTimeout(() => {
+      signalQmdProcessTree(child, "SIGKILL");
+      finish({
+        available: false,
+        reason: "binary",
+        error: `spawn ${params.command} timed out after ${timeoutMs}ms`,
+      });
+    }, timeoutMs);
+
+    child.once("error", (err) => {
+      finish({ available: false, reason: "binary", error: formatQmdAvailabilityError(err) });
+    });
+    child.once("spawn", () => {
+      didSpawn = true;
+      signalQmdProcessTree(child);
+      finish({ available: true });
+    });
+    child.once("close", () => {
+      if (!didSpawn) {
+        return;
+      }
+      finish({ available: true });
+    });
+  });
+}
+
+function validateQmdProbeCwd(cwd: string): QmdBinaryAvailability | null {
+  try {
+    const stat = statSync(cwd);
+    if (!stat.isDirectory()) {
+      return {
+        available: false,
+        reason: "workspace-cwd",
+        error: `workspace directory is not a directory: ${cwd}`,
+      };
+    }
+    return null;
+  } catch (err) {
+    if (typeof err === "object" && err && "code" in err && err.code === "ENOENT") {
+      return {
+        available: false,
+        reason: "workspace-cwd",
+        error: `workspace directory missing: ${cwd}`,
+      };
+    }
+    return {
+      available: false,
+      reason: "workspace-cwd",
+      error: `workspace directory unavailable: ${cwd} (${formatQmdAvailabilityError(err)})`,
+    };
+  }
+}
+
+/**
+ * Normalize an aborted signal into the error used to reject a killed command.
+ * Prefers the caller-supplied abort reason (so a deadline message survives) and
+ * falls back to a stable per-command abort error.
+ */
+function abortReason(signal: AbortSignal | undefined, commandSummary: string): Error {
+  const reason = signal?.reason;
+  if (reason instanceof Error) {
+    return reason;
+  }
+  if (typeof reason === "string" && reason.length > 0) {
+    return new Error(reason);
+  }
+  return new Error(`${commandSummary} aborted`);
+}
+
+export async function runCliCommand(params: {
+  commandSummary: string;
+  spawnInvocation: CliSpawnInvocation;
+  env: NodeJS.ProcessEnv;
+  cwd: string;
+  timeoutMs?: number;
+  maxOutputChars: number;
+  discardStdout?: boolean;
+  /**
+   * Caller-owned cancellation. When the signal aborts, the spawned child is
+   * killed immediately and the call rejects, so a caller that already stopped
+   * waiting (for example after its own deadline) does not leave an orphaned
+   * process running for the full command timeout.
+   */
+  signal?: AbortSignal;
+}): Promise<{ stdout: string; stderr: string }> {
+  return await new Promise((resolve, reject) => {
+    const { signal } = params;
+    if (signal?.aborted) {
+      reject(abortReason(signal, params.commandSummary));
+      return;
+    }
+    const child = spawn(params.spawnInvocation.command, params.spawnInvocation.argv, {
+      env: params.env,
+      cwd: params.cwd,
+      shell: params.spawnInvocation.shell,
+      windowsHide: params.spawnInvocation.windowsHide,
+      detached: shouldUseQmdProcessGroup(),
+    });
+    let stdout = "";
+    let stderr = "";
+    let stdoutTruncated = false;
+    let stderrTruncated = false;
+    let settled = false;
+    const discardStdout = params.discardStdout === true;
+    const timeoutMs =
+      params.timeoutMs === undefined ? undefined : resolveSafeTimeoutDelayMs(params.timeoutMs);
+    const timer = timeoutMs
+      ? setTimeout(() => {
+          signalQmdProcessTree(child, "SIGKILL");
+          settle(() =>
+            reject(new Error(`${params.commandSummary} timed out after ${timeoutMs}ms`)),
+          );
+        }, timeoutMs)
+      : null;
+    const onAbort = () => {
+      signalQmdProcessTree(child, "SIGKILL");
+      settle(() => reject(abortReason(signal, params.commandSummary)));
+    };
+    function settle(run: () => void): void {
+      if (settled) {
+        return;
+      }
+      settled = true;
+      if (timer) {
+        clearTimeout(timer);
+      }
+      signal?.removeEventListener("abort", onAbort);
+      run();
+    }
+    signal?.addEventListener("abort", onAbort, { once: true });
+    child.stdout.on("data", (data) => {
+      if (discardStdout) {
+        return;
+      }
+      const next = appendOutputWithCap(stdout, data.toString("utf8"), params.maxOutputChars);
+      stdout = next.text;
+      stdoutTruncated = stdoutTruncated || next.truncated;
+    });
+    child.stderr.on("data", (data) => {
+      const next = appendOutputWithCap(stderr, data.toString("utf8"), params.maxOutputChars);
+      stderr = next.text;
+      stderrTruncated = stderrTruncated || next.truncated;
+    });
+    child.on("error", (err) => {
+      if (timer) {
+        clearTimeout(timer);
+      }
+      settle(() => reject(err));
+    });
+    child.on("close", (code, closeSignal) => {
+      if (timer) {
+        clearTimeout(timer);
+      }
+      settle(() => {
+        if (!discardStdout && (stdoutTruncated || stderrTruncated)) {
+          reject(
+            new Error(
+              `${params.commandSummary} produced too much output (limit ${params.maxOutputChars} chars)`,
+            ),
+          );
+          return;
+        }
+        if (code === 0) {
+          resolve({ stdout, stderr });
+        } else {
+          reject(
+            new CliCommandError({
+              commandSummary: params.commandSummary,
+              code,
+              signal: closeSignal ?? null,
+              stdout,
+              stderr,
+            }),
+          );
+        }
+      });
+    });
+  });
+}
+
+function shouldUseQmdProcessGroup(): boolean {
+  return process.platform !== "win32";
+}
+
+function getEnvValueCaseInsensitive(
+  env: Record<string, string | undefined>,
+  expectedKey: string,
+): string | undefined {
+  const direct = env[expectedKey];
+  if (direct !== undefined) {
+    return direct;
+  }
+  const expected = expectedKey.toUpperCase();
+  const actualKey = Object.keys(env).find((key) => key.toUpperCase() === expected);
+  return actualKey ? env[actualKey] : undefined;
+}
+
+function normalizeWindowsSystemRoot(raw: string | undefined): string | null {
+  const trimmed = raw?.trim();
+  if (
+    !trimmed ||
+    trimmed.includes("\0") ||
+    trimmed.includes("\r") ||
+    trimmed.includes("\n") ||
+    trimmed.includes(";")
+  ) {
+    return null;
+  }
+  const normalized = path.win32.normalize(trimmed);
+  if (!path.win32.isAbsolute(normalized) || normalized.startsWith("\\\\")) {
+    return null;
+  }
+  const parsed = path.win32.parse(normalized);
+  if (!/^[A-Za-z]:\\$/.test(parsed.root) || normalized.length <= parsed.root.length) {
+    return null;
+  }
+  return normalized.replace(/[\\/]+$/, "");
+}
+
+function resolveWindowsTaskkillPath(env: Record<string, string | undefined> = process.env): string {
+  const systemRoot =
+    normalizeWindowsSystemRoot(getEnvValueCaseInsensitive(env, "SystemRoot")) ??
+    normalizeWindowsSystemRoot(getEnvValueCaseInsensitive(env, "WINDIR")) ??
+    DEFAULT_WINDOWS_SYSTEM_ROOT;
+  return path.win32.join(systemRoot, "System32", "taskkill.exe");
+}
+
+function signalQmdProcessTree(child: QmdChildProcess, signal?: NodeJS.Signals): void {
+  if (shouldUseQmdProcessGroup() && typeof child.pid === "number") {
+    try {
+      if (signal === undefined) {
+        process.kill(-child.pid);
+      } else {
+        process.kill(-child.pid, signal);
+      }
+      return;
+    } catch {
+      // Fall back to the direct child if the process group already disappeared.
+    }
+  }
+  if (!shouldUseQmdProcessGroup() && typeof child.pid === "number") {
+    const taskkillPath = resolveWindowsTaskkillPath();
+    const args = ["/PID", String(child.pid), "/T"];
+    if (signal === "SIGKILL") {
+      args.push("/F");
+    }
+    const result = spawnSync(taskkillPath, args, { stdio: "ignore", windowsHide: true });
+    if (!result.error && result.status === 0) {
+      return;
+    }
+    if (signal !== "SIGKILL") {
+      const forceResult = spawnSync(taskkillPath, [...args, "/F"], {
+        stdio: "ignore",
+        windowsHide: true,
+      });
+      if (!forceResult.error && forceResult.status === 0) {
+        return;
+      }
+    }
+  }
+  if (signal === undefined) {
+    child.kill();
+  } else {
+    child.kill(signal);
+  }
+}
+
+class CliCommandError extends Error {
+  readonly code: number | null;
+  readonly signal: NodeJS.Signals | null;
+  readonly stdout: string;
+  readonly stderr: string;
+
+  constructor(params: {
+    commandSummary: string;
+    code: number | null;
+    signal: NodeJS.Signals | null;
+    stdout: string;
+    stderr: string;
+  }) {
+    super(formatCliCommandFailureMessage(params));
+    this.name = "CliCommandError";
+    this.code = params.code;
+    this.signal = params.signal;
+    this.stdout = params.stdout;
+    this.stderr = params.stderr;
+  }
+}
+
+function formatCliCommandFailureMessage(params: {
+  commandSummary: string;
+  code: number | null;
+  signal: NodeJS.Signals | null;
+  stdout: string;
+  stderr: string;
+}): string {
+  const exit =
+    params.code === null ? `signal ${params.signal ?? "unknown"}` : `code ${String(params.code)}`;
+  return `${params.commandSummary} failed (${exit}): ${params.stderr || params.stdout}`;
+}
+
+function appendOutputWithCap(
+  current: string,
+  chunk: string,
+  maxChars: number,
+): { text: string; truncated: boolean } {
+  const appended = current + chunk;
+  const chars = Array.from(appended);
+  if (chars.length <= maxChars) {
+    return { text: appended, truncated: false };
+  }
+  return { text: chars.slice(-maxChars).join(""), truncated: true };
+}
+
+function formatQmdAvailabilityError(err: unknown): string {
+  if (err instanceof Error && err.message) {
+    return err.message;
+  }
+  return String(err);
+}

packages/memory-host-sdk/src/host/qmd-query-parser.test.ts

@@ -0,0 +1,74 @@
+// Memory Host SDK tests cover qmd query parser behavior.
+import { describe, expect, it } from "vitest";
+import { parseQmdQueryJson } from "./qmd-query-parser.js";
+
+describe("parseQmdQueryJson", () => {
+  it("parses clean qmd JSON output", () => {
+    const results = parseQmdQueryJson('[{"docid":"abc","score":1,"snippet":"@@ -1,1\\none"}]', "");
+    expect(results).toEqual([
+      {
+        docid: "abc",
+        score: 1,
+        snippet: "@@ -1,1\none",
+      },
+    ]);
+  });
+
+  it("extracts embedded result arrays from noisy stdout", () => {
+    const results = parseQmdQueryJson(
+      `initializing
+{"payload":"ok"}
+[{"docid":"abc","score":0.5}]
+complete`,
+      "",
+    );
+    expect(results).toEqual([{ docid: "abc", score: 0.5 }]);
+  });
+
+  it("preserves explicit qmd line metadata when present", () => {
+    const results = parseQmdQueryJson(
+      '[{"docid":"abc","score":0.5,"start_line":4,"end_line":6,"snippet":"@@ -10,1\\nignored"}]',
+      "",
+    );
+    expect(results).toEqual([
+      {
+        docid: "abc",
+        score: 0.5,
+        snippet: "@@ -10,1\nignored",
+        startLine: 4,
+        endLine: 6,
+      },
+    ]);
+  });
+
+  it("drops non-integer qmd line metadata", () => {
+    const results = parseQmdQueryJson(
+      `[{"docid":"abc","start_line":4.5,"end_line":${Number.MAX_SAFE_INTEGER + 1}}]`,
+      "",
+    );
+
+    expect(results).toEqual([{ docid: "abc" }]);
+  });
+
+  it("treats plain-text no-results from stderr as an empty result set", () => {
+    const results = parseQmdQueryJson("", "No results found\n");
+    expect(results).toStrictEqual([]);
+  });
+
+  it("treats prefixed no-results marker output as an empty result set", () => {
+    expect(parseQmdQueryJson("warning: no results found", "")).toStrictEqual([]);
+    expect(parseQmdQueryJson("", "[qmd] warning: no results found\n")).toStrictEqual([]);
+  });
+
+  it("does not treat arbitrary non-marker text as no-results output", () => {
+    expect(() =>
+      parseQmdQueryJson("warning: search completed; no results found for this query", ""),
+    ).toThrow(/qmd query returned invalid JSON/i);
+  });
+
+  it("throws when stdout cannot be interpreted as qmd JSON", () => {
+    expect(() => parseQmdQueryJson("this is not json", "")).toThrow(
+      /qmd query returned invalid JSON/i,
+    );
+  });
+});

packages/memory-host-sdk/src/host/qmd-query-parser.ts

@@ -0,0 +1,169 @@
+// Memory Host SDK module implements qmd query parser behavior.
+import { formatErrorMessage } from "./error-utils.js";
+import { normalizeLowercaseStringOrEmpty } from "./string-utils.js";
+
+// Parser for qmd query JSON output, including noisy CLI wrapper output.
+
+/** Normalized qmd query result consumed by memory search. */
+export type QmdQueryResult = {
+  docid?: string;
+  score?: number;
+  collection?: string;
+  file?: string;
+  snippet?: string;
+  body?: string;
+  startLine?: number;
+  endLine?: number;
+};
+
+/** Parse qmd stdout/stderr into normalized results, accepting known no-result markers. */
+export function parseQmdQueryJson(stdout: string, stderr: string): QmdQueryResult[] {
+  const trimmedStdout = stdout.trim();
+  const trimmedStderr = stderr.trim();
+  const stdoutIsMarker = trimmedStdout.length > 0 && isQmdNoResultsOutput(trimmedStdout);
+  const stderrIsMarker = trimmedStderr.length > 0 && isQmdNoResultsOutput(trimmedStderr);
+  if (stdoutIsMarker || (!trimmedStdout && stderrIsMarker)) {
+    return [];
+  }
+  if (!trimmedStdout) {
+    const context = trimmedStderr ? ` (stderr: ${summarizeQmdStderr(trimmedStderr)})` : "";
+    const message = `stdout empty${context}`;
+    warnQmdQueryParseError(message);
+    throw new Error(`qmd query returned invalid JSON: ${message}`);
+  }
+  try {
+    const parsed = parseQmdQueryResultArray(trimmedStdout);
+    if (parsed !== null) {
+      return parsed;
+    }
+    const noisyPayload = extractFirstJsonArray(trimmedStdout);
+    if (!noisyPayload) {
+      throw new Error("qmd query JSON response was not an array");
+    }
+    const fallback = parseQmdQueryResultArray(noisyPayload);
+    if (fallback !== null) {
+      return fallback;
+    }
+    throw new Error("qmd query JSON response was not an array");
+  } catch (err) {
+    const message = formatErrorMessage(err);
+    warnQmdQueryParseError(message);
+    throw new Error(`qmd query returned invalid JSON: ${message}`, { cause: err });
+  }
+}
+
+/** Emit parse warnings outside tests so broken qmd output is visible to operators. */
+function warnQmdQueryParseError(message: string): void {
+  if (process.env.VITEST || process.env.NODE_ENV === "test") {
+    return;
+  }
+  process.stderr.write(`qmd query returned invalid JSON: ${message}\n`);
+}
+
+/** Detect qmd no-result marker output on stdout or stderr. */
+function isQmdNoResultsOutput(raw: string): boolean {
+  const lines = raw
+    .split(/\r?\n/)
+    .map((line) => normalizeLowercaseStringOrEmpty(line).replace(/\s+/g, " "))
+    .filter((line) => line.length > 0);
+  return lines.some((line) => isQmdNoResultsLine(line));
+}
+
+/** Match qmd no-result lines with optional warning/info prefixes. */
+function isQmdNoResultsLine(line: string): boolean {
+  if (line === "no results found" || line === "no results found.") {
+    return true;
+  }
+  return /^(?:\[[^\]]+\]\s*)?(?:(?:warn(?:ing)?|info|error|qmd)\s*:\s*)+no results found\.?$/.test(
+    line,
+  );
+}
+
+/** Bound stderr context included in parse errors. */
+function summarizeQmdStderr(raw: string): string {
+  return raw.length <= 120 ? raw : `${raw.slice(0, 117)}...`;
+}
+
+/** Parse and normalize a strict qmd JSON array payload. */
+function parseQmdQueryResultArray(raw: string): QmdQueryResult[] | null {
+  try {
+    const parsed = JSON.parse(raw) as unknown;
+    if (!Array.isArray(parsed)) {
+      return null;
+    }
+    return parsed.map((item) => {
+      if (typeof item !== "object" || item === null) {
+        return item as QmdQueryResult;
+      }
+      const record = item as Record<string, unknown>;
+      const docid = typeof record.docid === "string" ? record.docid : undefined;
+      const score =
+        typeof record.score === "number" && Number.isFinite(record.score)
+          ? record.score
+          : undefined;
+      const collection = typeof record.collection === "string" ? record.collection : undefined;
+      const file = typeof record.file === "string" ? record.file : undefined;
+      const snippet = typeof record.snippet === "string" ? record.snippet : undefined;
+      const body = typeof record.body === "string" ? record.body : undefined;
+      return {
+        docid,
+        score,
+        collection,
+        file,
+        snippet,
+        body,
+        startLine: parseQmdLineNumber(record.start_line ?? record.startLine),
+        endLine: parseQmdLineNumber(record.end_line ?? record.endLine),
+      } as QmdQueryResult;
+    });
+  } catch {
+    return null;
+  }
+}
+
+/** Normalize qmd line numbers, rejecting zero, negative, and non-integer values. */
+function parseQmdLineNumber(value: unknown): number | undefined {
+  return typeof value === "number" && Number.isSafeInteger(value) && value > 0 ? value : undefined;
+}
+
+/** Extract the first complete JSON array from noisy stdout. */
+function extractFirstJsonArray(raw: string): string | null {
+  const start = raw.indexOf("[");
+  if (start < 0) {
+    return null;
+  }
+  let depth = 0;
+  let inString = false;
+  let escaped = false;
+  for (let i = start; i < raw.length; i += 1) {
+    const char = raw[i];
+    if (char === undefined) {
+      break;
+    }
+    if (inString) {
+      if (escaped) {
+        escaped = false;
+        continue;
+      }
+      if (char === "\\") {
+        escaped = true;
+      } else if (char === '"') {
+        inString = false;
+      }
+      continue;
+    }
+    if (char === '"') {
+      inString = true;
+      continue;
+    }
+    if (char === "[") {
+      depth += 1;
+    } else if (char === "]") {
+      depth -= 1;
+      if (depth === 0) {
+        return raw.slice(start, i + 1);
+      }
+    }
+  }
+  return null;
+}

packages/memory-host-sdk/src/host/qmd-scope.test.ts

@@ -0,0 +1,55 @@
+// Memory Host SDK tests cover qmd scope behavior.
+import { describe, expect, it } from "vitest";
+import type { ResolvedQmdConfig } from "./backend-config.js";
+import { deriveQmdScopeChannel, deriveQmdScopeChatType, isQmdScopeAllowed } from "./qmd-scope.js";
+
+describe("qmd scope", () => {
+  const allowDirect: ResolvedQmdConfig["scope"] = {
+    default: "deny",
+    rules: [{ action: "allow", match: { chatType: "direct" } }],
+  };
+
+  it("derives channel and chat type from canonical keys once", () => {
+    expect(deriveQmdScopeChannel("Workspace:group:123")).toBe("workspace");
+    expect(deriveQmdScopeChatType("Workspace:group:123")).toBe("group");
+  });
+
+  it("derives channel and chat type from stored key suffixes", () => {
+    expect(deriveQmdScopeChannel("agent:agent-1:workspace:channel:chan-123")).toBe("workspace");
+    expect(deriveQmdScopeChatType("agent:agent-1:workspace:channel:chan-123")).toBe("channel");
+  });
+
+  it("treats parsed keys with no chat prefix as direct", () => {
+    expect(deriveQmdScopeChannel("agent:agent-1:peer-direct")).toBeUndefined();
+    expect(deriveQmdScopeChatType("agent:agent-1:peer-direct")).toBe("direct");
+    expect(isQmdScopeAllowed(allowDirect, "agent:agent-1:peer-direct")).toBe(true);
+    expect(isQmdScopeAllowed(allowDirect, "agent:agent-1:peer:group:abc")).toBe(false);
+  });
+
+  it("applies scoped key-prefix checks against normalized key", () => {
+    const scope: ResolvedQmdConfig["scope"] = {
+      default: "deny",
+      rules: [{ action: "allow", match: { keyPrefix: "workspace:" } }],
+    };
+    expect(isQmdScopeAllowed(scope, "agent:agent-1:workspace:group:123")).toBe(true);
+    expect(isQmdScopeAllowed(scope, "agent:agent-1:other:group:123")).toBe(false);
+  });
+
+  it("supports rawKeyPrefix matches for agent-prefixed keys", () => {
+    const scope: ResolvedQmdConfig["scope"] = {
+      default: "allow",
+      rules: [{ action: "deny", match: { rawKeyPrefix: "agent:main:guildchat:" } }],
+    };
+    expect(isQmdScopeAllowed(scope, "agent:main:guildchat:channel:c123")).toBe(false);
+    expect(isQmdScopeAllowed(scope, "agent:main:workspace:channel:c123")).toBe(true);
+  });
+
+  it("keeps legacy agent-prefixed keyPrefix rules working", () => {
+    const scope: ResolvedQmdConfig["scope"] = {
+      default: "allow",
+      rules: [{ action: "deny", match: { keyPrefix: "agent:main:guildchat:" } }],
+    };
+    expect(isQmdScopeAllowed(scope, "agent:main:guildchat:channel:c123")).toBe(false);
+    expect(isQmdScopeAllowed(scope, "agent:main:workspace:channel:c123")).toBe(true);
+  });
+});

packages/memory-host-sdk/src/host/qmd-scope.ts

@@ -0,0 +1,123 @@
+// Memory Host SDK module implements qmd scope behavior.
+import type { ResolvedQmdConfig } from "./backend-config.js";
+import {
+  normalizeLowercaseStringOrEmpty,
+  normalizeOptionalLowercaseString,
+} from "./string-utils.js";
+
+type ParsedQmdSessionScope = {
+  channel?: string;
+  chatType?: "channel" | "group" | "direct";
+  normalizedKey?: string;
+};
+
+export function isQmdScopeAllowed(scope: ResolvedQmdConfig["scope"], sessionKey?: string): boolean {
+  if (!scope) {
+    return true;
+  }
+  const parsed = parseQmdSessionScope(sessionKey);
+  const channel = parsed.channel;
+  const chatType = parsed.chatType;
+  const normalizedKey = parsed.normalizedKey ?? "";
+  const rawKey = normalizeLowercaseStringOrEmpty(sessionKey ?? "");
+  for (const rule of scope.rules ?? []) {
+    if (!rule) {
+      continue;
+    }
+    const match = rule.match ?? {};
+    if (match.channel && match.channel !== channel) {
+      continue;
+    }
+    if (match.chatType && match.chatType !== chatType) {
+      continue;
+    }
+    const normalizedPrefix = normalizeOptionalLowercaseString(match.keyPrefix) || undefined;
+    const rawPrefix = normalizeOptionalLowercaseString(match.rawKeyPrefix) || undefined;
+
+    if (rawPrefix && !rawKey.startsWith(rawPrefix)) {
+      continue;
+    }
+    if (normalizedPrefix) {
+      // Backward compat: older configs used `keyPrefix: "agent:<id>:..."` to match raw keys.
+      const isLegacyRaw = normalizedPrefix.startsWith("agent:");
+      if (isLegacyRaw) {
+        if (!rawKey.startsWith(normalizedPrefix)) {
+          continue;
+        }
+      } else if (!normalizedKey.startsWith(normalizedPrefix)) {
+        continue;
+      }
+    }
+    return rule.action === "allow";
+  }
+  const fallback = scope.default ?? "allow";
+  return fallback === "allow";
+}
+
+export function deriveQmdScopeChannel(key?: string): string | undefined {
+  return parseQmdSessionScope(key).channel;
+}
+
+export function deriveQmdScopeChatType(key?: string): "channel" | "group" | "direct" | undefined {
+  return parseQmdSessionScope(key).chatType;
+}
+
+function parseQmdSessionScope(key?: string): ParsedQmdSessionScope {
+  const normalized = normalizeQmdSessionKey(key);
+  if (!normalized) {
+    return {};
+  }
+  const parts = normalized.split(":").filter(Boolean);
+  let chatType: ParsedQmdSessionScope["chatType"];
+  if (
+    parts.length >= 2 &&
+    (parts[1] === "group" || parts[1] === "channel" || parts[1] === "direct" || parts[1] === "dm")
+  ) {
+    if (parts.includes("group")) {
+      chatType = "group";
+    } else if (parts.includes("channel")) {
+      chatType = "channel";
+    }
+    return {
+      normalizedKey: normalized,
+      channel: normalizeOptionalLowercaseString(parts[0]),
+      chatType: chatType ?? "direct",
+    };
+  }
+  if (normalized.includes(":group:")) {
+    return { normalizedKey: normalized, chatType: "group" };
+  }
+  if (normalized.includes(":channel:")) {
+    return { normalizedKey: normalized, chatType: "channel" };
+  }
+  return { normalizedKey: normalized, chatType: "direct" };
+}
+
+function normalizeQmdSessionKey(key?: string): string | undefined {
+  if (!key) {
+    return undefined;
+  }
+  const trimmed = key.trim();
+  if (!trimmed) {
+    return undefined;
+  }
+  const parsed = parseAgentSessionKey(trimmed);
+  const normalized = normalizeLowercaseStringOrEmpty(parsed?.rest ?? trimmed);
+  if (normalized.startsWith("subagent:")) {
+    return undefined;
+  }
+  return normalized;
+}
+
+function parseAgentSessionKey(sessionKey: string | undefined | null): { rest: string } | null {
+  const raw = normalizeOptionalLowercaseString(sessionKey);
+  if (!raw) {
+    return null;
+  }
+  const parts = raw.split(":").filter(Boolean);
+  if (parts.length < 3 || parts[0] !== "agent") {
+    return null;
+  }
+  const rest = parts.slice(2).join(":");
+  return rest ? { rest } : null;
+}

packages/memory-host-sdk/src/host/query-expansion.test.ts

@@ -0,0 +1,184 @@
+// Memory Host SDK tests cover query keyword extraction behavior.
+import { describe, expect, it } from "vitest";
+import { extractKeywords } from "./query-expansion.js";
+
+describe("extractKeywords", () => {
+  it("extracts keywords from English conversational query", () => {
+    const keywords = extractKeywords("that thing we discussed about the API");
+    expect(keywords).toStrictEqual(["discussed", "api"]);
+  });
+
+  it("extracts keywords from Chinese conversational query", () => {
+    const keywords = extractKeywords("之前讨论的那个方案");
+    expect(keywords).toStrictEqual([
+      "之",
+      "讨",
+      "论",
+      "个",
+      "方",
+      "案",
+      "前讨",
+      "讨论",
+      "论的",
+      "的那",
+      "个方",
+      "方案",
+    ]);
+  });
+
+  it("extracts keywords from mixed language query", () => {
+    const keywords = extractKeywords("昨天讨论的 API design");
+    expect(keywords).toStrictEqual([
+      "昨",
+      "天",
+      "讨",
+      "论",
+      "天讨",
+      "讨论",
+      "论的",
+      "api",
+      "design",
+    ]);
+  });
+
+  it("returns specific technical terms", () => {
+    const keywords = extractKeywords("what was the solution for the CFR bug");
+    expect(keywords).toStrictEqual(["solution", "cfr", "bug"]);
+  });
+
+  it("extracts keywords from Korean conversational query", () => {
+    const keywords = extractKeywords("어제 논의한 배포 전략");
+    expect(keywords).toStrictEqual(["논의한", "배포", "전략"]);
+  });
+
+  it("strips Korean particles to extract stems", () => {
+    const keywords = extractKeywords("서버에서 발생한 에러를 확인");
+    expect(keywords).toStrictEqual(["서버에서", "서버", "발생한", "에러를", "에러", "확인"]);
+  });
+
+  it("filters Korean stop words including inflected forms", () => {
+    const keywords = extractKeywords("나는 그리고 그래서");
+    expect(keywords).toStrictEqual([]);
+  });
+
+  it("filters inflected Korean stop words not explicitly listed", () => {
+    const keywords = extractKeywords("그녀는 우리는");
+    expect(keywords).toStrictEqual([]);
+  });
+
+  it("does not produce bogus single-char stems from particle stripping", () => {
+    const keywords = extractKeywords("논의");
+    expect(keywords).toStrictEqual(["논의"]);
+  });
+
+  it("strips longest Korean trailing particles first", () => {
+    const keywords = extractKeywords("기능으로 설명");
+    expect(keywords).toStrictEqual(["기능으로", "기능", "설명"]);
+  });
+
+  it("keeps stripped ASCII stems for mixed Korean tokens", () => {
+    const keywords = extractKeywords("API를 배포했다");
+    expect(keywords).toStrictEqual(["api를", "api", "배포했다"]);
+  });
+
+  it("handles mixed Korean and English query", () => {
+    const keywords = extractKeywords("API 배포에 대한 논의");
+    expect(keywords).toStrictEqual(["api", "배포에", "배포", "대한", "논의"]);
+  });
+
+  it("extracts keywords from Japanese conversational query", () => {
+    const keywords = extractKeywords("昨日話したデプロイ戦略");
+    expect(keywords).toStrictEqual(["昨日話", "日話", "デプロイ", "戦略"]);
+  });
+
+  it("handles mixed Japanese and English query", () => {
+    const keywords = extractKeywords("昨日話したAPIのバグ");
+    expect(keywords).toStrictEqual(["昨日話", "日話", "api", "バグ"]);
+  });
+
+  it("filters Japanese stop words", () => {
+    const keywords = extractKeywords("これ それ そして どう");
+    expect(keywords).toStrictEqual([]);
+  });
+
+  it("extracts keywords from Spanish conversational query", () => {
+    const keywords = extractKeywords("ayer hablamos sobre la estrategia de despliegue");
+    expect(keywords).toStrictEqual(["hablamos", "estrategia", "despliegue"]);
+  });
+
+  it("extracts keywords from Portuguese conversational query", () => {
+    const keywords = extractKeywords("ontem falamos sobre a estratégia de implantação");
+    expect(keywords).toStrictEqual(["falamos", "estratégia", "implantação"]);
+  });
+
+  it("filters Spanish and Portuguese question stop words", () => {
+    const keywords = extractKeywords("cómo cuando donde porquê quando onde");
+    expect(keywords).toStrictEqual([]);
+  });
+
+  it("extracts keywords from Arabic conversational query", () => {
+    const keywords = extractKeywords("بالأمس ناقشنا استراتيجية النشر");
+    expect(keywords).toStrictEqual(["ناقشنا", "استراتيجية", "النشر"]);
+  });
+
+  it("filters Arabic question stop words", () => {
+    const keywords = extractKeywords("كيف متى أين ماذا");
+    expect(keywords).toStrictEqual([]);
+  });
+
+  it("handles empty query", () => {
+    expect(extractKeywords("")).toStrictEqual([]);
+    expect(extractKeywords("   ")).toStrictEqual([]);
+  });
+
+  it("handles query with only stop words", () => {
+    const keywords = extractKeywords("the a an is are");
+    expect(keywords).toStrictEqual([]);
+  });
+
+  it("removes duplicate keywords", () => {
+    const keywords = extractKeywords("test test testing");
+    expect(keywords).toStrictEqual(["test", "testing"]);
+  });
+
+  describe("with trigram tokenizer", () => {
+    const trigramOpts = { ftsTokenizer: "trigram" as const };
+
+    it("emits whole CJK block instead of unigrams in trigram mode", () => {
+      const defaultKeywords = extractKeywords("之前讨论的那个方案");
+      const trigramKeywords = extractKeywords("之前讨论的那个方案", trigramOpts);
+      expect(defaultKeywords).toStrictEqual([
+        "之",
+        "讨",
+        "论",
+        "个",
+        "方",
+        "案",
+        "前讨",
+        "讨论",
+        "论的",
+        "的那",
+        "个方",
+        "方案",
+      ]);
+      expect(trigramKeywords).toStrictEqual(["之前讨论的那个方案"]);
+    });
+
+    it("skips Japanese kanji bigrams in trigram mode", () => {
+      const defaultKeywords = extractKeywords("経済政策について");
+      const trigramKeywords = extractKeywords("経済政策について", trigramOpts);
+      expect(defaultKeywords).toStrictEqual(["経済政策", "経済", "済政", "政策", "について"]);
+      expect(trigramKeywords).toStrictEqual(["経済政策", "について"]);
+    });
+
+    it("still filters stop words in trigram mode", () => {
+      const keywords = extractKeywords("これ それ そして どう", trigramOpts);
+      expect(keywords).toStrictEqual([]);
+    });
+
+    it("does not affect English keyword extraction", () => {
+      const keywords = extractKeywords("that thing we discussed about the API", trigramOpts);
+      expect(keywords).toStrictEqual(["discussed", "api"]);
+    });
+  });
+});

packages/memory-host-sdk/src/host/query-expansion.ts

@@ -0,0 +1,776 @@
+// Memory Host SDK module implements query expansion behavior.
+import { normalizeLowercaseStringOrEmpty } from "./string-utils.js";
+
+/**
+ * Query expansion for FTS-only search mode.
+ *
+ * When no embedding provider is available, we fall back to FTS (full-text search).
+ * FTS works best with specific keywords, but users often ask conversational queries
+ * like "that thing we discussed yesterday" or "之前讨论的那个方案".
+ *
+ * This module extracts meaningful keywords from such queries to improve FTS results.
+ */
+
+// Common stop words that don't add search value
+const STOP_WORDS_EN = new Set([
+  // Articles and determiners
+  "a",
+  "an",
+  "the",
+  "this",
+  "that",
+  "these",
+  "those",
+  // Pronouns
+  "i",
+  "me",
+  "my",
+  "we",
+  "our",
+  "you",
+  "your",
+  "he",
+  "she",
+  "it",
+  "they",
+  "them",
+  // Common verbs
+  "is",
+  "are",
+  "was",
+  "were",
+  "be",
+  "been",
+  "being",
+  "have",
+  "has",
+  "had",
+  "do",
+  "does",
+  "did",
+  "will",
+  "would",
+  "could",
+  "should",
+  "can",
+  "may",
+  "might",
+  // Prepositions
+  "in",
+  "on",
+  "at",
+  "to",
+  "for",
+  "of",
+  "with",
+  "by",
+  "from",
+  "about",
+  "into",
+  "through",
+  "during",
+  "before",
+  "after",
+  "above",
+  "below",
+  "between",
+  "under",
+  "over",
+  // Conjunctions
+  "and",
+  "or",
+  "but",
+  "if",
+  "then",
+  "because",
+  "as",
+  "while",
+  "when",
+  "where",
+  "what",
+  "which",
+  "who",
+  "how",
+  "why",
+  // Time references (vague, not useful for FTS)
+  "yesterday",
+  "today",
+  "tomorrow",
+  "earlier",
+  "later",
+  "recently",
+  "before",
+  "ago",
+  "just",
+  "now",
+  // Vague references
+  "thing",
+  "things",
+  "stuff",
+  "something",
+  "anything",
+  "everything",
+  "nothing",
+  // Question words
+  "please",
+  "help",
+  "find",
+  "show",
+  "get",
+  "tell",
+  "give",
+]);
+
+const STOP_WORDS_ES = new Set([
+  // Articles and determiners
+  "el",
+  "la",
+  "los",
+  "las",
+  "un",
+  "una",
+  "unos",
+  "unas",
+  "este",
+  "esta",
+  "ese",
+  "esa",
+  // Pronouns
+  "yo",
+  "me",
+  "mi",
+  "nosotros",
+  "nosotras",
+  "tu",
+  "tus",
+  "usted",
+  "ustedes",
+  "ellos",
+  "ellas",
+  // Prepositions and conjunctions
+  "de",
+  "del",
+  "a",
+  "en",
+  "con",
+  "por",
+  "para",
+  "sobre",
+  "entre",
+  "y",
+  "o",
+  "pero",
+  "si",
+  "porque",
+  "como",
+  // Common verbs / auxiliaries
+  "es",
+  "son",
+  "fue",
+  "fueron",
+  "ser",
+  "estar",
+  "haber",
+  "tener",
+  "hacer",
+  // Time references (vague)
+  "ayer",
+  "hoy",
+  "mañana",
+  "antes",
+  "despues",
+  "después",
+  "ahora",
+  "recientemente",
+  // Question/request words
+  "que",
+  "qué",
+  "cómo",
+  "cuando",
+  "cuándo",
+  "donde",
+  "dónde",
+  "porqué",
+  "favor",
+  "ayuda",
+]);
+
+const STOP_WORDS_PT = new Set([
+  // Articles and determiners
+  "o",
+  "a",
+  "os",
+  "as",
+  "um",
+  "uma",
+  "uns",
+  "umas",
+  "este",
+  "esta",
+  "esse",
+  "essa",
+  // Pronouns
+  "eu",
+  "me",
+  "meu",
+  "minha",
+  "nos",
+  "nós",
+  "você",
+  "vocês",
+  "ele",
+  "ela",
+  "eles",
+  "elas",
+  // Prepositions and conjunctions
+  "de",
+  "do",
+  "da",
+  "em",
+  "com",
+  "por",
+  "para",
+  "sobre",
+  "entre",
+  "e",
+  "ou",
+  "mas",
+  "se",
+  "porque",
+  "como",
+  // Common verbs / auxiliaries
+  "é",
+  "são",
+  "foi",
+  "foram",
+  "ser",
+  "estar",
+  "ter",
+  "fazer",
+  // Time references (vague)
+  "ontem",
+  "hoje",
+  "amanhã",
+  "antes",
+  "depois",
+  "agora",
+  "recentemente",
+  // Question/request words
+  "que",
+  "quê",
+  "quando",
+  "onde",
+  "porquê",
+  "favor",
+  "ajuda",
+]);
+
+const STOP_WORDS_AR = new Set([
+  // Articles and connectors
+  "ال",
+  "و",
+  "أو",
+  "لكن",
+  "ثم",
+  "بل",
+  // Pronouns / references
+  "أنا",
+  "نحن",
+  "هو",
+  "هي",
+  "هم",
+  "هذا",
+  "هذه",
+  "ذلك",
+  "تلك",
+  "هنا",
+  "هناك",
+  // Common prepositions
+  "من",
+  "إلى",
+  "الى",
+  "في",
+  "على",
+  "عن",
+  "مع",
+  "بين",
+  "ل",
+  "ب",
+  "ك",
+  // Common auxiliaries / vague verbs
+  "كان",
+  "كانت",
+  "يكون",
+  "تكون",
+  "صار",
+  "أصبح",
+  "يمكن",
+  "ممكن",
+  // Time references (vague)
+  "بالأمس",
+  "امس",
+  "اليوم",
+  "غدا",
+  "الآن",
+  "قبل",
+  "بعد",
+  "مؤخرا",
+  // Question/request words
+  "لماذا",
+  "كيف",
+  "ماذا",
+  "متى",
+  "أين",
+  "هل",
+  "من فضلك",
+  "فضلا",
+  "ساعد",
+]);
+
+const STOP_WORDS_KO = new Set([
+  // Particles (조사)
+  "은",
+  "는",
+  "이",
+  "가",
+  "을",
+  "를",
+  "의",
+  "에",
+  "에서",
+  "로",
+  "으로",
+  "와",
+  "과",
+  "도",
+  "만",
+  "까지",
+  "부터",
+  "한테",
+  "에게",
+  "께",
+  "처럼",
+  "같이",
+  "보다",
+  "마다",
+  "밖에",
+  "대로",
+  // Pronouns (대명사)
+  "나",
+  "나는",
+  "내가",
+  "나를",
+  "너",
+  "우리",
+  "저",
+  "저희",
+  "그",
+  "그녀",
+  "그들",
+  "이것",
+  "저것",
+  "그것",
+  "여기",
+  "저기",
+  "거기",
+  // Common verbs / auxiliaries (일반 동사/보조 동사)
+  "있다",
+  "없다",
+  "하다",
+  "되다",
+  "이다",
+  "아니다",
+  "보다",
+  "주다",
+  "오다",
+  "가다",
+  // Nouns (의존 명사 / vague)
+  "것",
+  "거",
+  "등",
+  "수",
+  "때",
+  "곳",
+  "중",
+  "분",
+  // Adverbs
+  "잘",
+  "더",
+  "또",
+  "매우",
+  "정말",
+  "아주",
+  "많이",
+  "너무",
+  "좀",
+  // Conjunctions
+  "그리고",
+  "하지만",
+  "그래서",
+  "그런데",
+  "그러나",
+  "또는",
+  "그러면",
+  // Question words
+  "왜",
+  "어떻게",
+  "뭐",
+  "언제",
+  "어디",
+  "누구",
+  "무엇",
+  "어떤",
+  // Time (vague)
+  "어제",
+  "오늘",
+  "내일",
+  "최근",
+  "지금",
+  "아까",
+  "나중",
+  "전에",
+  // Request words
+  "제발",
+  "부탁",
+]);
+
+// Common Korean trailing particles to strip from words for tokenization
+// Sorted by descending length so longest-match-first is guaranteed.
+const KO_TRAILING_PARTICLES = [
+  "에서",
+  "으로",
+  "에게",
+  "한테",
+  "처럼",
+  "같이",
+  "보다",
+  "까지",
+  "부터",
+  "마다",
+  "밖에",
+  "대로",
+  "은",
+  "는",
+  "이",
+  "가",
+  "을",
+  "를",
+  "의",
+  "에",
+  "로",
+  "와",
+  "과",
+  "도",
+  "만",
+].toSorted((a, b) => b.length - a.length);
+
+function stripKoreanTrailingParticle(token: string): string | null {
+  for (const particle of KO_TRAILING_PARTICLES) {
+    if (token.length > particle.length && token.endsWith(particle)) {
+      return token.slice(0, -particle.length);
+    }
+  }
+  return null;
+}
+
+function isUsefulKoreanStem(stem: string): boolean {
+  // Prevent bogus one-syllable stems from words like "논의" -> "논".
+  if (/[\uac00-\ud7af]/.test(stem)) {
+    return stem.length >= 2;
+  }
+  // Keep stripped ASCII stems for mixed tokens like "API를" -> "api".
+  return /^[a-z0-9_]+$/i.test(stem);
+}
+
+const STOP_WORDS_JA = new Set([
+  // Pronouns and references
+  "これ",
+  "それ",
+  "あれ",
+  "この",
+  "その",
+  "あの",
+  "ここ",
+  "そこ",
+  "あそこ",
+  // Common auxiliaries / vague verbs
+  "する",
+  "した",
+  "して",
+  "です",
+  "ます",
+  "いる",
+  "ある",
+  "なる",
+  "できる",
+  // Particles / connectors
+  "の",
+  "こと",
+  "もの",
+  "ため",
+  "そして",
+  "しかし",
+  "また",
+  "でも",
+  "から",
+  "まで",
+  "より",
+  "だけ",
+  // Question words
+  "なぜ",
+  "どう",
+  "何",
+  "いつ",
+  "どこ",
+  "誰",
+  "どれ",
+  // Time (vague)
+  "昨日",
+  "今日",
+  "明日",
+  "最近",
+  "今",
+  "さっき",
+  "前",
+  "後",
+]);
+
+const STOP_WORDS_ZH = new Set([
+  // Pronouns
+  "我",
+  "我们",
+  "你",
+  "你们",
+  "他",
+  "她",
+  "它",
+  "他们",
+  "这",
+  "那",
+  "这个",
+  "那个",
+  "这些",
+  "那些",
+  // Auxiliary words
+  "的",
+  "了",
+  "着",
+  "过",
+  "得",
+  "地",
+  "吗",
+  "呢",
+  "吧",
+  "啊",
+  "呀",
+  "嘛",
+  "啦",
+  // Verbs (common, vague)
+  "是",
+  "有",
+  "在",
+  "被",
+  "把",
+  "给",
+  "让",
+  "用",
+  "到",
+  "去",
+  "来",
+  "做",
+  "说",
+  "看",
+  "找",
+  "想",
+  "要",
+  "能",
+  "会",
+  "可以",
+  // Prepositions and conjunctions
+  "和",
+  "与",
+  "或",
+  "但",
+  "但是",
+  "因为",
+  "所以",
+  "如果",
+  "虽然",
+  "而",
+  "也",
+  "都",
+  "就",
+  "还",
+  "又",
+  "再",
+  "才",
+  "只",
+  // Time (vague)
+  "之前",
+  "以前",
+  "之后",
+  "以后",
+  "刚才",
+  "现在",
+  "昨天",
+  "今天",
+  "明天",
+  "最近",
+  // Vague references
+  "东西",
+  "事情",
+  "事",
+  "什么",
+  "哪个",
+  "哪些",
+  "怎么",
+  "为什么",
+  "多少",
+  // Question/request words
+  "请",
+  "帮",
+  "帮忙",
+  "告诉",
+]);
+
+/** Returns true for low-value conversational tokens that should not drive FTS matching. */
+export function isQueryStopWordToken(token: string): boolean {
+  return (
+    STOP_WORDS_EN.has(token) ||
+    STOP_WORDS_ES.has(token) ||
+    STOP_WORDS_PT.has(token) ||
+    STOP_WORDS_AR.has(token) ||
+    STOP_WORDS_ZH.has(token) ||
+    STOP_WORDS_KO.has(token) ||
+    STOP_WORDS_JA.has(token)
+  );
+}
+
+/**
+ * Check if a token looks like a meaningful keyword.
+ * Returns false for short tokens, numbers-only, etc.
+ */
+function isValidKeyword(token: string): boolean {
+  if (!token || token.length === 0) {
+    return false;
+  }
+  // Skip very short English words (likely stop words or fragments)
+  if (/^[a-zA-Z]+$/.test(token) && token.length < 3) {
+    return false;
+  }
+  // Skip pure numbers (not useful for semantic search)
+  if (/^\d+$/.test(token)) {
+    return false;
+  }
+  // Skip tokens that are all punctuation
+  if (/^[\p{P}\p{S}]+$/u.test(token)) {
+    return false;
+  }
+  return true;
+}
+
+/**
+ * Simple tokenizer that handles English, Chinese, Korean, and Japanese text.
+ * For Chinese, we do character-based splitting since we don't have a proper segmenter.
+ * For English, we split on whitespace and punctuation.
+ */
+function tokenize(text: string, opts?: { ftsTokenizer?: "unicode61" | "trigram" }): string[] {
+  const useTrigram = opts?.ftsTokenizer === "trigram";
+  const tokens: string[] = [];
+  const normalized = normalizeLowercaseStringOrEmpty(text);
+
+  // Split into segments (English words, Chinese character sequences, etc.)
+  const segments = normalized.split(/[\s\p{P}]+/u).filter(Boolean);
+
+  for (const segment of segments) {
+    // Japanese text often mixes scripts (kanji/kana/ASCII) without spaces.
+    // Extract script-specific chunks so technical terms like "API" / "バグ" are retained.
+    if (/[\u3040-\u30ff]/.test(segment)) {
+      const jpParts =
+        segment.match(/[a-z0-9_]+|[\u30a0-\u30ffー]+|[\u4e00-\u9fff]+|[\u3040-\u309f]{2,}/g) ?? [];
+      for (const part of jpParts) {
+        if (/^[\u4e00-\u9fff]+$/.test(part)) {
+          tokens.push(part);
+          if (!useTrigram) {
+            for (let i = 0; i < part.length - 1; i++) {
+              tokens.push(part[i] + part[i + 1]);
+            }
+          }
+        } else {
+          tokens.push(part);
+        }
+      }
+    } else if (/[\u4e00-\u9fff]/.test(segment)) {
+      // Check if segment contains CJK characters (Chinese)
+      const chars = Array.from(segment).filter((c) => /[\u4e00-\u9fff]/.test(c));
+      if (useTrigram) {
+        // In trigram mode, push the whole contiguous CJK block (mirroring the
+        // Japanese kanji path). SQLite's trigram FTS requires at least 3 characters
+        // per query term — individual characters silently return no results.
+        const block = chars.join("");
+        if (block.length > 0) {
+          tokens.push(block);
+        }
+      } else {
+        // Default mode: unigrams + bigrams for phrase matching
+        tokens.push(...chars);
+        for (let i = 0; i < chars.length - 1; i++) {
+          tokens.push(chars[i] + chars[i + 1]);
+        }
+      }
+    } else if (/[\uac00-\ud7af\u3131-\u3163]/.test(segment)) {
+      // For Korean (Hangul syllables and jamo), keep the word as-is unless it is
+      // effectively a stop word once trailing particles are removed.
+      const stem = stripKoreanTrailingParticle(segment);
+      const stemIsStopWord = stem !== null && STOP_WORDS_KO.has(stem);
+      if (!STOP_WORDS_KO.has(segment) && !stemIsStopWord) {
+        tokens.push(segment);
+      }
+      // Also emit particle-stripped stems when they are useful keywords.
+      if (stem && !STOP_WORDS_KO.has(stem) && isUsefulKoreanStem(stem)) {
+        tokens.push(stem);
+      }
+    } else {
+      // For non-CJK, keep as single token
+      tokens.push(segment);
+    }
+  }
+
+  return tokens;
+}
+
+/**
+ * Extract keywords from a conversational query for FTS search.
+ *
+ * Examples:
+ * - "that thing we discussed about the API" → ["discussed", "API"]
+ * - "之前讨论的那个方案" → ["讨论", "方案"]
+ * - "what was the solution for the bug" → ["solution", "bug"]
+ */
+export function extractKeywords(
+  query: string,
+  opts?: { ftsTokenizer?: "unicode61" | "trigram" },
+): string[] {
+  const tokens = tokenize(query, opts);
+  const keywords: string[] = [];
+  const seen = new Set<string>();
+
+  for (const token of tokens) {
+    // Skip stop words
+    if (isQueryStopWordToken(token)) {
+      continue;
+    }
+    // Skip invalid keywords
+    if (!isValidKeyword(token)) {
+      continue;
+    }
+    // Skip duplicates
+    if (seen.has(token)) {
+      continue;
+    }
+    seen.add(token);
+    keywords.push(token);
+  }
+
+  return keywords;
+}

packages/memory-host-sdk/src/host/read-file-shared.test.ts

@@ -0,0 +1,37 @@
+// Memory Host SDK tests cover read file shared behavior.
+import { describe, expect, it } from "vitest";
+import { buildMemoryReadResult, buildMemoryReadResultFromSlice } from "./read-file-shared.js";
+
+describe("memory read result slicing", () => {
+  it("uses default line windows for non-finite from and lines values", () => {
+    expect(
+      buildMemoryReadResult({
+        content: "one\ntwo\nthree",
+        relPath: "memory/test.md",
+        from: Number.NaN,
+        lines: Number.NaN,
+      }),
+    ).toEqual({
+      text: "one\ntwo\nthree",
+      path: "memory/test.md",
+      from: 1,
+      lines: 3,
+    });
+  });
+
+  it("uses the default character budget for non-finite maxChars values", () => {
+    expect(
+      buildMemoryReadResultFromSlice({
+        selectedLines: ["one", "two"],
+        relPath: "memory/test.md",
+        startLine: Number.POSITIVE_INFINITY,
+        maxChars: Number.NaN,
+      }),
+    ).toEqual({
+      text: "one\ntwo",
+      path: "memory/test.md",
+      from: 1,
+      lines: 2,
+    });
+  });
+});

packages/memory-host-sdk/src/host/read-file-shared.ts

@@ -0,0 +1,130 @@
+// Memory Host SDK module implements read file shared behavior.
+import type { MemoryReadResult } from "./types.js";
+
+// Shared memory-file read result shaping and truncation notices.
+
+/** Default number of lines returned by memory read helpers. */
+export const DEFAULT_MEMORY_READ_LINES = 120;
+/** Default max character budget for memory read helper output. */
+export const DEFAULT_MEMORY_READ_MAX_CHARS = 12_000;
+
+export type { MemoryReadResult } from "./types.js";
+
+/** Build the continuation notice appended to truncated memory excerpts. */
+function buildContinuationNotice(params: {
+  nextFrom: number | undefined;
+  suggestReadFallback?: boolean;
+}): string {
+  const base =
+    typeof params.nextFrom === "number"
+      ? `[More content available. Use from=${params.nextFrom} to continue.]`
+      : "[More content available. Requested excerpt exceeded the default maxChars budget.]";
+  const fallback = params.suggestReadFallback
+    ? " If you need the full raw line, use read on the source file."
+    : "";
+  return `\n\n${base.slice(0, -1)}${fallback}]`;
+}
+
+/** Fit line slices to the response character budget while preserving line boundaries. */
+function fitLinesToCharBudget(params: { lines: string[]; maxChars: number }): {
+  text: string;
+  includedLines: number;
+  hardTruncatedSingleLine: boolean;
+} {
+  const { lines, maxChars } = params;
+  if (lines.length === 0) {
+    return { text: "", includedLines: 0, hardTruncatedSingleLine: false };
+  }
+
+  let includedLines = lines.length;
+  let text = lines.join("\n");
+  while (includedLines > 1 && text.length > maxChars) {
+    includedLines -= 1;
+    text = lines.slice(0, includedLines).join("\n");
+  }
+
+  if (text.length <= maxChars) {
+    return { text, includedLines, hardTruncatedSingleLine: false };
+  }
+
+  return {
+    text: text.slice(0, maxChars),
+    includedLines: 1,
+    hardTruncatedSingleLine: true,
+  };
+}
+
+/** Normalize optional numeric config to a positive integer fallback. */
+function normalizePositiveInteger(value: number | undefined, fallback: number): number {
+  return typeof value === "number" && Number.isFinite(value)
+    ? Math.max(1, Math.floor(value))
+    : fallback;
+}
+
+/** Build a memory read result from an already-selected line slice. */
+export function buildMemoryReadResultFromSlice(params: {
+  selectedLines: string[];
+  relPath: string;
+  startLine: number;
+  moreSourceLinesRemain?: boolean;
+  maxChars?: number;
+  suggestReadFallback?: boolean;
+}): MemoryReadResult {
+  const start = normalizePositiveInteger(params.startLine, 1);
+  const fitted = fitLinesToCharBudget({
+    lines: params.selectedLines,
+    maxChars: normalizePositiveInteger(params.maxChars, DEFAULT_MEMORY_READ_MAX_CHARS),
+  });
+  const moreSourceLinesRemain = params.moreSourceLinesRemain ?? false;
+  const charCapTruncated =
+    fitted.hardTruncatedSingleLine || fitted.includedLines < params.selectedLines.length;
+  const nextFrom =
+    !fitted.hardTruncatedSingleLine &&
+    (moreSourceLinesRemain || fitted.includedLines < params.selectedLines.length)
+      ? start + fitted.includedLines
+      : undefined;
+  const truncated = charCapTruncated || moreSourceLinesRemain;
+  const text =
+    truncated && fitted.text
+      ? `${fitted.text}${buildContinuationNotice({
+          nextFrom,
+          suggestReadFallback: fitted.hardTruncatedSingleLine && params.suggestReadFallback,
+        })}`
+      : fitted.text;
+  return {
+    text,
+    path: params.relPath,
+    from: start,
+    lines: fitted.includedLines,
+    ...(truncated ? { truncated: true } : {}),
+    ...(typeof nextFrom === "number" ? { nextFrom } : {}),
+  };
+}
+
+/** Build a memory read result from raw file content and caller range options. */
+export function buildMemoryReadResult(params: {
+  content: string;
+  relPath: string;
+  from?: number;
+  lines?: number;
+  defaultLines?: number;
+  maxChars?: number;
+  suggestReadFallback?: boolean;
+}): MemoryReadResult {
+  const fileLines = params.content.split("\n");
+  const start = normalizePositiveInteger(params.from, 1);
+  const requestedCount = normalizePositiveInteger(
+    params.lines ?? params.defaultLines,
+    DEFAULT_MEMORY_READ_LINES,
+  );
+  const selectedLines = fileLines.slice(start - 1, start - 1 + requestedCount);
+  const moreSourceLinesRemain = start - 1 + selectedLines.length < fileLines.length;
+  return buildMemoryReadResultFromSlice({
+    selectedLines,
+    relPath: params.relPath,
+    startLine: start,
+    moreSourceLinesRemain,
+    maxChars: params.maxChars,
+    suggestReadFallback: params.suggestReadFallback,
+  });
+}

packages/memory-host-sdk/src/host/read-file.test.ts

@@ -0,0 +1,161 @@
+// Memory Host SDK tests cover read file behavior.
+import fs from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { describe, expect, it, vi } from "vitest";
+import { readMemoryFile } from "./read-file.js";
+
+async function createDirectorySymlink(target: string, linkPath: string): Promise<boolean> {
+  try {
+    await fs.symlink(target, linkPath, "dir");
+    return true;
+  } catch (err) {
+    const code = (err as NodeJS.ErrnoException).code;
+    if (code === "EPERM" || code === "EACCES") {
+      return false;
+    }
+    throw err;
+  }
+}
+
+describe("readMemoryFile", () => {
+  it("returns empty text for missing files under extra path directories", async () => {
+    const tmpRoot = await fs.mkdtemp(path.join(os.tmpdir(), "memory-read-file-"));
+    try {
+      const workspaceDir = path.join(tmpRoot, "workspace");
+      const extraDir = path.join(tmpRoot, "extra");
+      const missingPath = path.join(extraDir, "missing.md");
+      await fs.mkdir(workspaceDir, { recursive: true });
+      await fs.mkdir(extraDir, { recursive: true });
+
+      const result = await readMemoryFile({
+        workspaceDir,
+        extraPaths: [extraDir],
+        relPath: missingPath,
+      });
+
+      expect(result).toEqual({
+        text: "",
+        path: path.relative(workspaceDir, missingPath).replace(/\\/g, "/"),
+      });
+
+      const nonDirectoryParentPath = path.join(extraDir, "note.md", "child.md");
+      await fs.writeFile(path.join(extraDir, "note.md"), "note", "utf-8");
+      await expect(
+        readMemoryFile({
+          workspaceDir,
+          extraPaths: [extraDir],
+          relPath: nonDirectoryParentPath,
+        }),
+      ).resolves.toEqual({
+        text: "",
+        path: path.relative(workspaceDir, nonDirectoryParentPath).replace(/\\/g, "/"),
+      });
+    } finally {
+      await fs.rm(tmpRoot, { recursive: true, force: true });
+    }
+  });
+
+  it("rejects extra path reads through symlinked directory components", async () => {
+    const tmpRoot = await fs.mkdtemp(path.join(os.tmpdir(), "memory-read-file-"));
+    try {
+      const workspaceDir = path.join(tmpRoot, "workspace");
+      const extraDir = path.join(tmpRoot, "extra");
+      const outsideDir = path.join(tmpRoot, "outside");
+      await fs.mkdir(workspaceDir, { recursive: true });
+      await fs.mkdir(extraDir, { recursive: true });
+      await fs.mkdir(outsideDir, { recursive: true });
+      await fs.writeFile(path.join(extraDir, "inside.md"), "inside", "utf-8");
+      await fs.writeFile(path.join(outsideDir, "private.md"), "private", "utf-8");
+
+      const inside = await readMemoryFile({
+        workspaceDir,
+        extraPaths: [extraDir],
+        relPath: path.join(extraDir, "inside.md"),
+      });
+      expect(inside.text).toBe("inside");
+
+      const insideLinkPath = path.join(extraDir, "inside-link");
+      if (!(await createDirectorySymlink(extraDir, insideLinkPath))) {
+        return;
+      }
+      await expect(
+        readMemoryFile({
+          workspaceDir,
+          extraPaths: [extraDir],
+          relPath: path.join(insideLinkPath, "inside.md"),
+        }),
+      ).rejects.toThrow("path required");
+
+      const outsideLinkPath = path.join(extraDir, "link");
+      if (!(await createDirectorySymlink(outsideDir, outsideLinkPath))) {
+        return;
+      }
+
+      await expect(
+        readMemoryFile({
+          workspaceDir,
+          extraPaths: [extraDir],
+          relPath: path.join(outsideLinkPath, "private.md"),
+        }),
+      ).rejects.toThrow("path required");
+      await expect(
+        readMemoryFile({
+          workspaceDir,
+          extraPaths: [extraDir],
+          relPath: path.join(outsideLinkPath, "missing.md"),
+        }),
+      ).rejects.toThrow("path required");
+    } finally {
+      await fs.rm(tmpRoot, { recursive: true, force: true });
+    }
+  });
+
+  it("retries transient read errors for workspace memory files", async () => {
+    const tmpRoot = await fs.mkdtemp(path.join(os.tmpdir(), "memory-read-file-"));
+    try {
+      const workspaceDir = path.join(tmpRoot, "workspace");
+      const relPath = "memory/retry.md";
+      const absPath = path.join(workspaceDir, relPath);
+      await fs.mkdir(path.dirname(absPath), { recursive: true });
+      await fs.writeFile(absPath, "alpha\nbeta", "utf-8");
+
+      const realOpen = fs.open;
+      let attempts = 0;
+      const openSpy = vi
+        .spyOn(fs, "open")
+        .mockImplementation(async (...args: Parameters<typeof realOpen>) => {
+          const [target, flags, mode] = args;
+          if (typeof target === "string" && path.resolve(target) === absPath && attempts++ === 0) {
+            const err = new Error(
+              "Unknown system error -11: Unknown system error -11, open",
+            ) as NodeJS.ErrnoException;
+            err.code = "UNKNOWN";
+            err.errno = -11;
+            throw err;
+          }
+          return await realOpen(target, flags, mode);
+        });
+
+      try {
+        await expect(
+          readMemoryFile({
+            workspaceDir,
+            extraPaths: [],
+            relPath,
+          }),
+        ).resolves.toEqual({
+          text: "alpha\nbeta",
+          path: relPath,
+          from: 1,
+          lines: 2,
+        });
+        expect(attempts).toBe(2);
+      } finally {
+        openSpy.mockRestore();
+      }
+    } finally {
+      await fs.rm(tmpRoot, { recursive: true, force: true });
+    }
+  });
+});

packages/memory-host-sdk/src/host/read-file.ts

@@ -0,0 +1,182 @@
+// Memory Host SDK module implements read file behavior.
+import fs from "node:fs/promises";
+import path from "node:path";
+import {
+  resolveAgentContextLimits,
+  resolveAgentWorkspaceDir,
+  resolveMemorySearchConfig,
+  type OpenClawConfig,
+} from "./config-utils.js";
+import {
+  assertNoSymlinkParents,
+  isFileMissingError,
+  isPathInside,
+  isPathInsideWithRealpath,
+  readRegularFile,
+  root,
+  statRegularFile,
+} from "./fs-utils.js";
+import { isMemoryPath, normalizeExtraMemoryPaths } from "./internal.js";
+import {
+  buildMemoryReadResult,
+  DEFAULT_MEMORY_READ_LINES,
+  type MemoryReadResult,
+} from "./read-file-shared.js";
+import { retryTransientMemoryRead } from "./read-retry.js";
+
+// Secure markdown memory-file reader for workspace and configured extra paths.
+
+/** Check that an absolute path stays inside an allowed extra directory without symlink escapes. */
+async function isAllowedAdditionalDirectoryPath(
+  additionalPath: string,
+  absPath: string,
+): Promise<boolean> {
+  if (!isPathInside(additionalPath, absPath)) {
+    return false;
+  }
+  try {
+    await assertNoSymlinkParents({ rootDir: additionalPath, targetPath: absPath });
+  } catch {
+    return false;
+  }
+  if (!isPathInsideWithRealpath(additionalPath, absPath)) {
+    try {
+      await fs.lstat(absPath);
+    } catch (err) {
+      return isFileMissingError(err);
+    }
+    return false;
+  }
+  return true;
+}
+
+/** Return true when a file vanished after path validation but before content read. */
+function isFileDisappearedDuringReadError(err: unknown): boolean {
+  return (
+    isFileMissingError(err) ||
+    Boolean(
+      err &&
+      typeof err === "object" &&
+      "code" in err &&
+      (err as { code?: unknown }).code === "path-mismatch",
+    )
+  );
+}
+
+/** Read a validated memory markdown file from workspace or configured extra paths. */
+export async function readMemoryFile(params: {
+  workspaceDir: string;
+  extraPaths?: string[];
+  relPath: string;
+  from?: number;
+  lines?: number;
+  defaultLines?: number;
+  maxChars?: number;
+}): Promise<MemoryReadResult> {
+  const rawPath = params.relPath.trim();
+  if (!rawPath) {
+    throw new Error("path required");
+  }
+  const absPath = path.isAbsolute(rawPath)
+    ? path.resolve(rawPath)
+    : path.resolve(params.workspaceDir, rawPath);
+  const relPath = path.relative(params.workspaceDir, absPath).replace(/\\/g, "/");
+  const inWorkspace = relPath.length > 0 && !relPath.startsWith("..") && !path.isAbsolute(relPath);
+  const allowedWorkspace = inWorkspace && isMemoryPath(relPath);
+  let allowedAdditional = false;
+  if (!allowedWorkspace && (params.extraPaths?.length ?? 0) > 0) {
+    const additionalPaths = normalizeExtraMemoryPaths(params.workspaceDir, params.extraPaths);
+    for (const additionalPath of additionalPaths) {
+      try {
+        const stat = await fs.lstat(additionalPath);
+        if (stat.isSymbolicLink()) {
+          continue;
+        }
+        if (stat.isDirectory()) {
+          if (await isAllowedAdditionalDirectoryPath(additionalPath, absPath)) {
+            const candidateStat = await fs.lstat(absPath).catch(() => null);
+            if (candidateStat?.isSymbolicLink()) {
+              continue;
+            }
+            allowedAdditional = true;
+            break;
+          }
+          continue;
+        }
+        if (stat.isFile() && absPath === additionalPath && absPath.endsWith(".md")) {
+          allowedAdditional = true;
+          break;
+        }
+      } catch {}
+    }
+  }
+  if (!allowedWorkspace && !allowedAdditional) {
+    throw new Error("path required");
+  }
+  if (!absPath.endsWith(".md")) {
+    throw new Error("path required");
+  }
+  if (allowedWorkspace) {
+    try {
+      // Workspace reads use the safe fs root so symlink escapes are rejected before file IO.
+      const workspaceRoot = await root(params.workspaceDir);
+      await workspaceRoot.resolve(relPath);
+    } catch (err) {
+      if (isFileMissingError(err)) {
+        return { text: "", path: relPath };
+      }
+      throw err;
+    }
+  }
+  const statResult = await statRegularFile(absPath);
+  if (statResult.missing) {
+    return { text: "", path: relPath };
+  }
+  let content: string;
+  try {
+    content = (
+      await retryTransientMemoryRead(
+        () => readRegularFile({ filePath: absPath }),
+        `read memory file ${absPath}`,
+      )
+    ).buffer.toString("utf-8");
+  } catch (err) {
+    if (isFileDisappearedDuringReadError(err)) {
+      return { text: "", path: relPath };
+    }
+    throw err;
+  }
+  return buildMemoryReadResult({
+    content,
+    relPath,
+    from: params.from,
+    lines: params.lines,
+    defaultLines: params.defaultLines ?? DEFAULT_MEMORY_READ_LINES,
+    maxChars: params.maxChars,
+    suggestReadFallback: allowedWorkspace,
+  });
+}
+
+/** Resolve agent memory config and read one memory file for that agent. */
+export async function readAgentMemoryFile(params: {
+  cfg: OpenClawConfig;
+  agentId: string;
+  relPath: string;
+  from?: number;
+  lines?: number;
+}): Promise<MemoryReadResult> {
+  const settings = resolveMemorySearchConfig(params.cfg, params.agentId);
+  if (!settings) {
+    throw new Error("memory search disabled");
+  }
+  const contextLimits = resolveAgentContextLimits(params.cfg, params.agentId);
+  return await readMemoryFile({
+    workspaceDir: resolveAgentWorkspaceDir(params.cfg, params.agentId),
+    extraPaths: settings.extraPaths,
+    relPath: params.relPath,
+    from: params.from,
+    lines: params.lines,
+    defaultLines: contextLimits?.memoryGetDefaultLines,
+    maxChars: contextLimits?.memoryGetMaxChars,
+  });
+}

packages/memory-host-sdk/src/host/read-retry.test.ts

@@ -0,0 +1,26 @@
+// Memory Host SDK tests cover read retry behavior.
+import { afterEach, describe, expect, it, vi } from "vitest";
+import { retryTransientMemoryRead } from "./read-retry.js";
+
+afterEach(() => {
+  vi.restoreAllMocks();
+});
+
+describe("retryTransientMemoryRead", () => {
+  it("uses a short two-retry budget for transient file read errors", async () => {
+    const err = new Error("Unknown system error -11: Unknown system error -11, read");
+    const run = vi.fn<() => Promise<string>>().mockRejectedValue(err);
+    const timeoutSpy = vi.spyOn(globalThis, "setTimeout").mockImplementation((callback) => {
+      if (typeof callback === "function") {
+        callback();
+      }
+      return 0 as unknown as ReturnType<typeof setTimeout>;
+    });
+
+    await expect(retryTransientMemoryRead(run)).rejects.toThrow("Unknown system error -11");
+
+    expect(run).toHaveBeenCalledTimes(3);
+    expect(timeoutSpy).toHaveBeenNthCalledWith(1, expect.any(Function), 25);
+    expect(timeoutSpy).toHaveBeenNthCalledWith(2, expect.any(Function), 50);
+  });
+});

packages/memory-host-sdk/src/host/read-retry.ts

@@ -0,0 +1,51 @@
+// Memory Host SDK module implements read retry behavior.
+import { retryAsync } from "./retry-utils.js";
+
+// Retry helper for transient filesystem reads observed on memory stores.
+
+const TRANSIENT_MEMORY_READ_ERRNO = -11;
+const TRANSIENT_MEMORY_READ_CODES = new Set(["EAGAIN", "EWOULDBLOCK", "EDEADLK"]);
+const TRANSIENT_MEMORY_READ_MESSAGE = /Unknown system error -11\b/i;
+
+/** Extract errno from Node filesystem-style errors. */
+function getErrno(error: unknown): number | undefined {
+  return typeof (error as NodeJS.ErrnoException | undefined)?.errno === "number"
+    ? (error as NodeJS.ErrnoException).errno
+    : undefined;
+}
+
+/** Extract code from Node filesystem-style errors. */
+function getCode(error: unknown): string | undefined {
+  return typeof (error as NodeJS.ErrnoException | undefined)?.code === "string"
+    ? (error as NodeJS.ErrnoException).code
+    : undefined;
+}
+
+/** Return true for transient memory read failures that should be retried. */
+export function isTransientMemoryReadError(error: unknown): boolean {
+  const code = getCode(error);
+  if (code && TRANSIENT_MEMORY_READ_CODES.has(code)) {
+    return true;
+  }
+
+  const errno = getErrno(error);
+  if (errno === TRANSIENT_MEMORY_READ_ERRNO) {
+    return true;
+  }
+
+  return error instanceof Error && TRANSIENT_MEMORY_READ_MESSAGE.test(error.message);
+}
+
+/** Retry a memory read with the narrow transient error predicate. */
+export async function retryTransientMemoryRead<T>(
+  read: () => Promise<T>,
+  label = "memory read",
+): Promise<T> {
+  return await retryAsync(read, {
+    attempts: 3,
+    minDelayMs: 25,
+    maxDelayMs: 50,
+    label,
+    shouldRetry: (error) => isTransientMemoryReadError(error),
+  });
+}

packages/memory-host-sdk/src/host/remote-http.test.ts

@@ -0,0 +1,61 @@
+// Memory Host SDK tests cover remote http behavior.
+import { describe, expect, it } from "vitest";
+import { MEMORY_REMOTE_TRUSTED_ENV_PROXY_MODE, withRemoteHttpResponse } from "./remote-http.js";
+
+describe("package withRemoteHttpResponse", () => {
+  function makeFetchDeps({ useEnvProxy = false }: { useEnvProxy?: boolean } = {}) {
+    const calls: unknown[] = [];
+    return {
+      calls,
+      fetchWithSsrFGuardImpl: async (params: unknown) => {
+        calls.push(params);
+        return {
+          response: new Response("ok", { status: 200 }),
+          finalUrl: "https://memory.example/v1",
+          release: async () => {},
+        };
+      },
+      shouldUseEnvHttpProxyForUrlImpl: () => useEnvProxy,
+    };
+  }
+
+  it("uses trusted env proxy mode when the target will use EnvHttpProxyAgent", async () => {
+    const deps = makeFetchDeps({ useEnvProxy: true });
+
+    await withRemoteHttpResponse({
+      url: "https://memory.example/v1/embeddings",
+      onResponse: async () => undefined,
+      ...deps,
+    });
+
+    expect(deps.calls[0]).toHaveProperty("url", "https://memory.example/v1/embeddings");
+    expect(deps.calls[0]).toHaveProperty("mode", MEMORY_REMOTE_TRUSTED_ENV_PROXY_MODE);
+  });
+
+  it("keeps strict guarded fetch mode when proxy env would not proxy the target", async () => {
+    const deps = makeFetchDeps();
+
+    await withRemoteHttpResponse({
+      url: "https://internal.corp.example/v1/embeddings",
+      onResponse: async () => undefined,
+      ...deps,
+    });
+
+    expect(deps.calls).toHaveLength(1);
+    expect(deps.calls[0]).not.toHaveProperty("mode");
+  });
+
+  it("passes abort signals to the guarded fetch", async () => {
+    const deps = makeFetchDeps();
+    const controller = new AbortController();
+
+    await withRemoteHttpResponse({
+      url: "https://memory.example/v1/embeddings",
+      signal: controller.signal,
+      onResponse: async () => undefined,
+      ...deps,
+    });
+
+    expect(deps.calls[0]).toHaveProperty("signal", controller.signal);
+  });
+});

packages/memory-host-sdk/src/host/remote-http.ts

@@ -0,0 +1,46 @@
+// Memory Host SDK module implements remote http behavior.
+import {
+  fetchWithSsrFGuard,
+  shouldUseEnvHttpProxyForUrl,
+  ssrfPolicyFromHttpBaseUrlAllowedHostname,
+} from "./openclaw-runtime-network.js";
+import type { SsrFPolicy } from "./ssrf-policy.js";
+
+// Remote memory HTTP wrapper that applies SSRF policy and releases guarded sockets.
+
+/** Proxy mode used only for URLs that the runtime classified as env-proxy safe. */
+export const MEMORY_REMOTE_TRUSTED_ENV_PROXY_MODE = "trusted_env_proxy";
+
+/** Build an SSRF allow policy from a configured remote base URL. */
+export const buildRemoteBaseUrlPolicy: (baseUrl: string) => SsrFPolicy | undefined =
+  ssrfPolicyFromHttpBaseUrlAllowedHostname;
+
+/** Execute a remote HTTP request under SSRF guard and always release the response handle. */
+export async function withRemoteHttpResponse<T>(params: {
+  url: string;
+  init?: RequestInit;
+  signal?: AbortSignal;
+  ssrfPolicy?: SsrFPolicy;
+  fetchImpl?: typeof fetch;
+  fetchWithSsrFGuardImpl?: typeof fetchWithSsrFGuard;
+  shouldUseEnvHttpProxyForUrlImpl?: typeof shouldUseEnvHttpProxyForUrl;
+  auditContext?: string;
+  onResponse: (response: Response) => Promise<T>;
+}): Promise<T> {
+  const guardedFetch = params.fetchWithSsrFGuardImpl ?? fetchWithSsrFGuard;
+  const shouldUseEnvProxy = params.shouldUseEnvHttpProxyForUrlImpl ?? shouldUseEnvHttpProxyForUrl;
+  const { response, release } = await guardedFetch({
+    url: params.url,
+    fetchImpl: params.fetchImpl,
+    init: params.init,
+    signal: params.signal,
+    policy: params.ssrfPolicy,
+    auditContext: params.auditContext ?? "memory-remote",
+    ...(shouldUseEnvProxy(params.url) ? { mode: MEMORY_REMOTE_TRUSTED_ENV_PROXY_MODE } : {}),
+  });
+  try {
+    return await params.onResponse(response);
+  } finally {
+    await release();
+  }
+}

packages/memory-host-sdk/src/host/response-snippet.test.ts

@@ -0,0 +1,77 @@
+// Memory Host SDK tests cover response snippet behavior.
+import { describe, expect, it } from "vitest";
+import { readResponseJsonWithLimit, readResponseTextSnippet } from "./response-snippet.js";
+
+describe("readResponseTextSnippet", () => {
+  function stallingResponse(onCancel: () => void): Response {
+    const reader = {
+      read: () => new Promise<ReadableStreamReadResult<Uint8Array>>(() => {}),
+      cancel: async () => {
+        onCancel();
+      },
+      releaseLock: () => undefined,
+    } as ReadableStreamDefaultReader<Uint8Array>;
+
+    return {
+      body: { getReader: () => reader },
+      headers: new Headers(),
+    } as Response;
+  }
+
+  it("does not wait for another chunk after reading the byte cap exactly", async () => {
+    let canceled = false;
+    const stream = new ReadableStream<Uint8Array>({
+      start(controller) {
+        controller.enqueue(new TextEncoder().encode("abcd"));
+      },
+      cancel() {
+        canceled = true;
+      },
+    });
+
+    await expect(
+      readResponseTextSnippet(new Response(stream), { maxBytes: 4, maxChars: 100 }),
+    ).resolves.toBe("abcd... [truncated]");
+    expect(canceled).toBe(true);
+  });
+
+  it("cancels snippet body reads when the caller signal aborts", async () => {
+    let canceled = false;
+    const response = stallingResponse(() => {
+      canceled = true;
+    });
+    const controller = new AbortController();
+    const read = readResponseTextSnippet(response, {
+      maxBytes: 1024,
+      signal: controller.signal,
+    });
+
+    await new Promise((resolve) => {
+      setTimeout(resolve, 0);
+    });
+    controller.abort(new Error("snippet aborted"));
+
+    await expect(read).rejects.toThrow("snippet aborted");
+    expect(canceled).toBe(true);
+  });
+
+  it("cancels JSON body reads when the caller signal aborts", async () => {
+    let canceled = false;
+    const response = stallingResponse(() => {
+      canceled = true;
+    });
+    const controller = new AbortController();
+    const read = readResponseJsonWithLimit(response, {
+      errorPrefix: "remote memory",
+      signal: controller.signal,
+    });
+
+    await new Promise((resolve) => {
+      setTimeout(resolve, 0);
+    });
+    controller.abort(new Error("json aborted"));
+
+    await expect(read).rejects.toThrow("json aborted");
+    expect(canceled).toBe(true);
+  });
+});

packages/memory-host-sdk/src/host/response-snippet.ts

@@ -0,0 +1,247 @@
+// Memory Host SDK module implements response snippet behavior.
+const DEFAULT_ERROR_BODY_MAX_BYTES = 8 * 1024;
+const DEFAULT_ERROR_BODY_MAX_CHARS = 1_000;
+const DEFAULT_JSON_BODY_MAX_BYTES = 64 * 1024 * 1024;
+const TRUNCATED_SUFFIX = "... [truncated]";
+
+// Bounded response readers for provider/remote HTTP errors and JSON bodies.
+type ResponseTextSnippetOptions = {
+  maxBytes?: number;
+  maxChars?: number;
+  signal?: AbortSignal;
+};
+
+type ResponseJsonOptions = {
+  maxBytes?: number;
+  errorPrefix: string;
+  signal?: AbortSignal;
+};
+
+type ResponsePrefix = {
+  bytes: Uint8Array[];
+  length: number;
+  truncated: boolean;
+};
+
+/** Read a small collapsed text snippet from a response body. */
+export async function readResponseTextSnippet(
+  res: Response,
+  options: ResponseTextSnippetOptions = {},
+): Promise<string> {
+  const maxBytes = options.maxBytes ?? DEFAULT_ERROR_BODY_MAX_BYTES;
+  const maxChars = options.maxChars ?? DEFAULT_ERROR_BODY_MAX_CHARS;
+  const prefix = await readResponsePrefix(res, maxBytes, options.signal);
+  if (prefix.length === 0) {
+    return "";
+  }
+
+  const text = new TextDecoder().decode(joinChunks(prefix.bytes, prefix.length));
+  const collapsed = text.replace(/\s+/g, " ").trim();
+  if (!collapsed) {
+    return "";
+  }
+  if (prefix.truncated || collapsed.length > maxChars) {
+    return `${collapsed.slice(0, maxChars)}${TRUNCATED_SUFFIX}`;
+  }
+  return collapsed;
+}
+
+/** Read and parse JSON while enforcing a hard byte limit. */
+export async function readResponseJsonWithLimit(
+  res: Response,
+  options: ResponseJsonOptions,
+): Promise<unknown> {
+  const maxBytes = options.maxBytes ?? DEFAULT_JSON_BODY_MAX_BYTES;
+  const contentLength = parseContentLength(res.headers.get("content-length"), options.errorPrefix);
+  if (typeof contentLength === "number" && contentLength > maxBytes) {
+    await cancelResponseBody(res);
+    throw responseTooLarge(options.errorPrefix, contentLength, maxBytes);
+  }
+
+  const text = await readResponseTextWithLimit(res, maxBytes, options.errorPrefix, options.signal);
+
+  try {
+    return JSON.parse(text);
+  } catch (cause) {
+    throw new Error(`${options.errorPrefix}: malformed JSON response`, { cause });
+  }
+}
+
+function toAbortError(signal: AbortSignal, fallbackMessage: string): Error {
+  return signal.reason instanceof Error ? signal.reason : new Error(fallbackMessage);
+}
+
+async function readChunkWithAbort(
+  reader: ReadableStreamDefaultReader<Uint8Array>,
+  signal: AbortSignal | undefined,
+  fallbackMessage: string,
+): Promise<ReadableStreamReadResult<Uint8Array>> {
+  if (!signal) {
+    return await reader.read();
+  }
+  if (signal.aborted) {
+    await reader.cancel().catch(() => undefined);
+    throw toAbortError(signal, fallbackMessage);
+  }
+
+  let removeAbortListener: (() => void) | undefined;
+  const abortPromise = new Promise<ReadableStreamReadResult<Uint8Array>>((_resolve, reject) => {
+    const onAbort = () => {
+      void reader.cancel().catch(() => undefined);
+      reject(toAbortError(signal, fallbackMessage));
+    };
+    signal.addEventListener("abort", onAbort, { once: true });
+    removeAbortListener = () => signal.removeEventListener("abort", onAbort);
+  });
+
+  try {
+    return await Promise.race([reader.read(), abortPromise]);
+  } finally {
+    removeAbortListener?.();
+  }
+}
+
+async function readResponsePrefix(
+  res: Response,
+  maxBytes: number,
+  signal?: AbortSignal,
+): Promise<ResponsePrefix> {
+  const body = res.body;
+  if (!body || typeof body.getReader !== "function") {
+    return { bytes: [], length: 0, truncated: false };
+  }
+
+  const reader = body.getReader();
+  const chunks: Uint8Array[] = [];
+  let length = 0;
+  let truncated = false;
+
+  try {
+    while (true) {
+      const { done, value } = await readChunkWithAbort(
+        reader,
+        signal,
+        "Response snippet body read aborted",
+      );
+      if (done) {
+        break;
+      }
+      if (!value?.length) {
+        continue;
+      }
+
+      const remaining = maxBytes - length;
+      if (value.length >= remaining) {
+        // Keep only the configured prefix and cancel the body so callers do not
+        // accidentally buffer large provider error responses.
+        if (remaining > 0) {
+          chunks.push(value.subarray(0, remaining));
+          length += remaining;
+        }
+        truncated = true;
+        await reader.cancel().catch(() => undefined);
+        break;
+      }
+
+      chunks.push(value);
+      length += value.length;
+    }
+  } finally {
+    try {
+      reader.releaseLock();
+    } catch {}
+  }
+
+  return { bytes: chunks, length, truncated };
+}
+
+async function readResponseTextWithLimit(
+  res: Response,
+  maxBytes: number,
+  errorPrefix: string,
+  signal?: AbortSignal,
+): Promise<string> {
+  const body = res.body;
+  if (!body || typeof body.getReader !== "function") {
+    return "";
+  }
+
+  const reader = body.getReader();
+  const chunks: Uint8Array[] = [];
+  let length = 0;
+
+  try {
+    while (true) {
+      const { done, value } = await readChunkWithAbort(
+        reader,
+        signal,
+        `${errorPrefix}: response body read aborted`,
+      );
+      if (done) {
+        break;
+      }
+      if (!value?.length) {
+        continue;
+      }
+
+      const nextLength = length + value.length;
+      if (nextLength > maxBytes) {
+        await reader.cancel().catch(() => undefined);
+        throw responseTooLarge(errorPrefix, nextLength, maxBytes);
+      }
+
+      chunks.push(value);
+      length = nextLength;
+    }
+  } finally {
+    try {
+      reader.releaseLock();
+    } catch {}
+  }
+
+  return new TextDecoder().decode(joinChunks(chunks, length));
+}
+
+async function cancelResponseBody(res: Response): Promise<void> {
+  const body = res.body;
+  if (!body || typeof body.cancel !== "function") {
+    return;
+  }
+  await body.cancel().catch(() => undefined);
+}
+
+function parseContentLength(raw: string | null, errorPrefix: string): number | undefined {
+  const trimmed = raw?.trim();
+  if (!trimmed) {
+    return undefined;
+  }
+  if (!/^(0|[1-9]\d*)$/.test(trimmed)) {
+    throw new Error(`${errorPrefix}: invalid content-length header: ${raw}`);
+  }
+  const value = Number(trimmed);
+  if (!Number.isSafeInteger(value)) {
+    throw new Error(`${errorPrefix}: invalid content-length header: ${raw}`);
+  }
+  return value;
+}
+
+function responseTooLarge(errorPrefix: string, size: number, maxBytes: number): Error {
+  return new Error(responseTooLargeMessage(errorPrefix, size, maxBytes));
+}
+
+function responseTooLargeMessage(errorPrefix: string, size: number, maxBytes: number): string {
+  return `${errorPrefix}: response body too large: ${size} bytes (limit: ${maxBytes} bytes)`;
+}
+
+function joinChunks(chunks: Uint8Array[], length: number): Uint8Array {
+  if (chunks.length === 1 && chunks[0]?.length === length) {
+    return chunks[0];
+  }
+  const joined = new Uint8Array(length);
+  let offset = 0;
+  for (const chunk of chunks) {
+    joined.set(chunk, offset);
+    offset += chunk.length;
+  }
+  return joined;
+}

packages/memory-host-sdk/src/host/retry-utils.test.ts

@@ -0,0 +1,86 @@
+// Memory Host SDK tests cover retry utils behavior.
+import { afterEach, describe, expect, it, vi } from "vitest";
+import { MAX_SAFE_TIMEOUT_DELAY_MS } from "../../../gateway-client/src/timeouts.js";
+import { resolveRetryConfig, retryAsync } from "./retry-utils.js";
+
+afterEach(() => {
+  vi.restoreAllMocks();
+});
+
+describe("resolveRetryConfig", () => {
+  const defaults = {
+    attempts: 4,
+    minDelayMs: 0,
+    maxDelayMs: 0,
+    jitter: 0,
+  };
+
+  it("does not round malformed attempt counts", () => {
+    expect(resolveRetryConfig(defaults, { attempts: 1.5 }).attempts).toBe(4);
+    expect(resolveRetryConfig(defaults, { attempts: Number.POSITIVE_INFINITY }).attempts).toBe(4);
+    expect(resolveRetryConfig(defaults, { attempts: Number.NaN }).attempts).toBe(4);
+  });
+
+  it("caps oversized retry delays at the timer-safe ceiling", () => {
+    const config = resolveRetryConfig(defaults, {
+      minDelayMs: Number.MAX_SAFE_INTEGER,
+      maxDelayMs: Number.MAX_SAFE_INTEGER,
+    });
+
+    expect(config.minDelayMs).toBe(MAX_SAFE_TIMEOUT_DELAY_MS);
+    expect(config.maxDelayMs).toBe(MAX_SAFE_TIMEOUT_DELAY_MS);
+  });
+});
+
+describe("retryAsync", () => {
+  it("falls back to the default attempt count for malformed numeric counts", async () => {
+    const run = vi.fn(async () => {
+      throw new Error("boom");
+    });
+
+    await expect(retryAsync(run, Number.NaN, 0)).rejects.toThrow("boom");
+
+    expect(run).toHaveBeenCalledTimes(3);
+  });
+
+  it("caps legacy numeric retry sleeps at the timer-safe ceiling", async () => {
+    const run = vi
+      .fn<() => Promise<string>>()
+      .mockRejectedValueOnce(new Error("boom"))
+      .mockResolvedValueOnce("ok");
+    const timeoutSpy = vi.spyOn(globalThis, "setTimeout").mockImplementation((callback) => {
+      if (typeof callback === "function") {
+        callback();
+      }
+      return 0 as unknown as ReturnType<typeof setTimeout>;
+    });
+
+    await expect(retryAsync(run, 2, Number.MAX_SAFE_INTEGER)).resolves.toBe("ok");
+
+    expect(timeoutSpy).toHaveBeenCalledWith(expect.any(Function), MAX_SAFE_TIMEOUT_DELAY_MS);
+  });
+
+  it("caps retryAfterMs sleeps at the timer-safe ceiling", async () => {
+    const run = vi
+      .fn<() => Promise<string>>()
+      .mockRejectedValueOnce(new Error("boom"))
+      .mockResolvedValueOnce("ok");
+    const timeoutSpy = vi.spyOn(globalThis, "setTimeout").mockImplementation((callback) => {
+      if (typeof callback === "function") {
+        callback();
+      }
+      return 0 as unknown as ReturnType<typeof setTimeout>;
+    });
+
+    await expect(
+      retryAsync(run, {
+        attempts: 2,
+        minDelayMs: 0,
+        maxDelayMs: Number.MAX_SAFE_INTEGER,
+        retryAfterMs: () => Number.MAX_SAFE_INTEGER,
+      }),
+    ).resolves.toBe("ok");
+
+    expect(timeoutSpy).toHaveBeenCalledWith(expect.any(Function), MAX_SAFE_TIMEOUT_DELAY_MS);
+  });
+});

packages/memory-host-sdk/src/host/retry-utils.ts

@@ -0,0 +1,175 @@
+// Memory Host SDK helper module supports retry utils behavior.
+import { resolveSafeTimeoutDelayMs } from "../../../gateway-client/src/timeouts.js";
+
+/** Retry timing configuration with optional jitter. */
+export type RetryConfig = {
+  attempts?: number;
+  minDelayMs?: number;
+  maxDelayMs?: number;
+  jitter?: number;
+};
+
+/** Retry callback payload. */
+export type RetryInfo = {
+  attempt: number;
+  maxAttempts: number;
+  delayMs: number;
+  err: unknown;
+  label?: string;
+};
+
+/** Retry options for retryAsync. */
+export type RetryOptions = RetryConfig & {
+  label?: string;
+  shouldRetry?: (err: unknown, attempt: number) => boolean;
+  retryAfterMs?: (err: unknown) => number | undefined;
+  onRetry?: (info: RetryInfo) => void;
+};
+
+const DEFAULT_RETRY_CONFIG = {
+  attempts: 3,
+  minDelayMs: 300,
+  maxDelayMs: 30_000,
+  jitter: 0,
+};
+
+function sleep(ms: number): Promise<void> {
+  return new Promise((resolve) => {
+    setTimeout(resolve, ms);
+  });
+}
+
+function asFiniteNumber(value: unknown): number | undefined {
+  if (typeof value !== "number" || !Number.isFinite(value)) {
+    return undefined;
+  }
+  return value;
+}
+
+function clampNumber(value: unknown, fallback: number, min?: number, max?: number): number {
+  const next = asFiniteNumber(value);
+  if (next === undefined) {
+    return fallback;
+  }
+  const floor = typeof min === "number" ? min : Number.NEGATIVE_INFINITY;
+  const ceiling = typeof max === "number" ? max : Number.POSITIVE_INFINITY;
+  return Math.min(Math.max(next, floor), ceiling);
+}
+
+function resolveAttempts(value: unknown, fallback: number): number {
+  if (typeof value !== "number" || !Number.isSafeInteger(value)) {
+    return fallback;
+  }
+  return Math.max(1, value);
+}
+
+/** Resolve retry settings with clamped positive timeout values. */
+export function resolveRetryConfig(
+  defaults: Required<RetryConfig> = DEFAULT_RETRY_CONFIG,
+  overrides?: RetryConfig,
+): Required<RetryConfig> {
+  const attempts = resolveAttempts(overrides?.attempts, defaults.attempts);
+  const minDelayMs = resolveSafeTimeoutDelayMs(
+    Math.round(clampNumber(overrides?.minDelayMs, defaults.minDelayMs, 0)),
+    { minMs: 0 },
+  );
+  const maxDelayMs = Math.max(
+    minDelayMs,
+    resolveSafeTimeoutDelayMs(
+      Math.round(clampNumber(overrides?.maxDelayMs, defaults.maxDelayMs, 0)),
+      { minMs: 0 },
+    ),
+  );
+  const jitter = clampNumber(overrides?.jitter, defaults.jitter, 0, 1);
+  return { attempts, minDelayMs, maxDelayMs, jitter };
+}
+
+function applyJitter(delayMs: number, jitter: number): number {
+  if (jitter <= 0) {
+    return delayMs;
+  }
+  const offset = (Math.random() * 2 - 1) * jitter;
+  return Math.max(0, Math.round(delayMs * (1 + offset)));
+}
+
+/** Run an async operation with exponential backoff retry handling. */
+export async function retryAsync<T>(
+  fn: () => Promise<T>,
+  attemptsOrOptions: number | RetryOptions = 3,
+  initialDelayMs = 300,
+): Promise<T> {
+  if (typeof attemptsOrOptions === "number") {
+    const attempts = resolveAttempts(attemptsOrOptions, DEFAULT_RETRY_CONFIG.attempts);
+    let lastErr: unknown;
+    for (let i = 0; i < attempts; i += 1) {
+      try {
+        return await fn();
+      } catch (err) {
+        lastErr = err;
+        if (i === attempts - 1) {
+          break;
+        }
+        await sleep(resolveSafeTimeoutDelayMs(initialDelayMs * 2 ** i, { minMs: 0 }));
+      }
+    }
+    throw toLintErrorObject(lastErr ?? new Error("Retry failed"), "Non-Error thrown");
+  }
+
+  const options = attemptsOrOptions;
+  const resolved = resolveRetryConfig(DEFAULT_RETRY_CONFIG, options);
+  const maxAttempts = resolved.attempts;
+  const minDelayMs = resolved.minDelayMs;
+  const maxDelayMs =
+    Number.isFinite(resolved.maxDelayMs) && resolved.maxDelayMs > 0
+      ? resolved.maxDelayMs
+      : Number.POSITIVE_INFINITY;
+  const shouldRetry = options.shouldRetry ?? (() => true);
+  let lastErr: unknown;
+
+  for (let attempt = 1; attempt <= maxAttempts; attempt += 1) {
+    try {
+      return await fn();
+    } catch (err) {
+      lastErr = err;
+      if (attempt >= maxAttempts || !shouldRetry(err, attempt)) {
+        break;
+      }
+
+      const retryAfterMs = options.retryAfterMs?.(err);
+      const hasRetryAfter = typeof retryAfterMs === "number" && Number.isFinite(retryAfterMs);
+      const baseDelay = hasRetryAfter
+        ? Math.max(resolveSafeTimeoutDelayMs(retryAfterMs, { minMs: 0 }), minDelayMs)
+        : resolveSafeTimeoutDelayMs(minDelayMs * 2 ** (attempt - 1), { minMs: 0 });
+      let delay = Math.min(baseDelay, maxDelayMs);
+      delay = applyJitter(delay, resolved.jitter);
+      delay = Math.min(Math.max(delay, minDelayMs), maxDelayMs);
+
+      options.onRetry?.({
+        attempt,
+        maxAttempts,
+        delayMs: delay,
+        err,
+        label: options.label,
+      });
+      if (delay > 0) {
+        await sleep(delay);
+      }
+    }
+  }
+
+  throw toLintErrorObject(lastErr ?? new Error("Retry failed"), "Non-Error thrown");
+}
+
+export function toLintErrorObject(value: unknown, fallbackMessage: string): Error {
+  if (value instanceof Error) {
+    return value;
+  }
+  if (typeof value === "string") {
+    return new Error(value);
+  }
+  const error = new Error(fallbackMessage, { cause: value });
+  if ((typeof value === "object" && value !== null) || typeof value === "function") {
+    Object.assign(error, value);
+  }
+  return error;
+}

packages/memory-host-sdk/src/host/secret-input-utils.ts

@@ -0,0 +1,164 @@
+// Secret input parsing shared by memory provider config and gateway-resolved snapshots.
+
+/** Supported secret reference backing stores. */
+export type SecretRefSource = "env" | "file" | "exec";
+
+/** Canonical secret reference shape used after gateway resolution. */
+export type SecretRef = {
+  source: SecretRefSource;
+  provider: string;
+  id: string;
+};
+
+const DEFAULT_SECRET_PROVIDER_ALIAS = "default";
+const ENV_SECRET_REF_ID_RE = /^[A-Z][A-Z0-9_]{0,127}$/;
+const LEGACY_SECRETREF_ENV_MARKER_PREFIX = "secretref-env:";
+const ENV_SECRET_TEMPLATE_RE = /^\$\{([A-Z][A-Z0-9_]{0,127})\}$/;
+const SECRET_REF_SOURCES = new Set<SecretRefSource>(["env", "file", "exec"]);
+
+/** Narrow unknown JSON config values to plain records. */
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+
+/** Normalize literal secret strings and reject empty placeholders. */
+function normalizeSecretInputString(value: unknown): string | undefined {
+  if (typeof value !== "string") {
+    return undefined;
+  }
+  const trimmed = value.trim();
+  return trimmed.length > 0 ? trimmed : undefined;
+}
+
+/** Narrow a string to a supported SecretRef source. */
+function hasSecretRefSource(value: unknown): value is SecretRefSource {
+  return typeof value === "string" && SECRET_REF_SOURCES.has(value as SecretRefSource);
+}
+
+/** Narrow unknown values to non-empty strings. */
+function hasNonEmptyString(value: unknown): value is string {
+  return typeof value === "string" && value.trim().length > 0;
+}
+
+/** Detect canonical three-field SecretRef objects. */
+function isSecretRef(value: unknown): value is SecretRef {
+  if (!isRecord(value)) {
+    return false;
+  }
+  const keys = Object.keys(value);
+  return (
+    keys.length === 3 &&
+    hasSecretRefSource(value.source) &&
+    hasNonEmptyString(value.provider) &&
+    hasNonEmptyString(value.id)
+  );
+}
+
+/** Detect legacy refs that predate explicit provider names. */
+function isLegacySecretRefWithoutProvider(
+  value: unknown,
+): value is { source: SecretRefSource; id: string } {
+  if (!isRecord(value)) {
+    return false;
+  }
+  return (
+    hasSecretRefSource(value.source) && hasNonEmptyString(value.id) && value.provider === undefined
+  );
+}
+
+/** Parse env template shorthand such as "${OPENAI_API_KEY}". */
+function parseEnvTemplateSecretRef(value: unknown): SecretRef | null {
+  if (typeof value !== "string") {
+    return null;
+  }
+  const match = ENV_SECRET_TEMPLATE_RE.exec(value.trim());
+  if (!match) {
+    return null;
+  }
+  return {
+    source: "env",
+    provider: DEFAULT_SECRET_PROVIDER_ALIAS,
+    id: match[1] ?? "",
+  };
+}
+
+/** Parse legacy secretref-env markers from older config snapshots. */
+function parseLegacySecretRefEnvMarker(value: unknown): SecretRef | null {
+  if (typeof value !== "string") {
+    return null;
+  }
+  const trimmed = value.trim();
+  if (!trimmed.startsWith(LEGACY_SECRETREF_ENV_MARKER_PREFIX)) {
+    return null;
+  }
+  const id = trimmed.slice(LEGACY_SECRETREF_ENV_MARKER_PREFIX.length);
+  if (!ENV_SECRET_REF_ID_RE.test(id)) {
+    return null;
+  }
+  return {
+    source: "env",
+    provider: DEFAULT_SECRET_PROVIDER_ALIAS,
+    id,
+  };
+}
+
+/** Coerce all accepted shipped secret reference shapes to canonical SecretRef. */
+function coerceSecretRef(value: unknown): SecretRef | null {
+  if (isSecretRef(value)) {
+    return value;
+  }
+  if (isLegacySecretRefWithoutProvider(value)) {
+    return {
+      source: value.source,
+      provider: DEFAULT_SECRET_PROVIDER_ALIAS,
+      id: value.id,
+    };
+  }
+  return parseEnvTemplateSecretRef(value) ?? parseLegacySecretRefEnvMarker(value);
+}
+
+/** Return true when a secret input has either a literal value or resolvable reference shape. */
+export function hasConfiguredSecretInput(value: unknown): boolean {
+  if (normalizeSecretInputString(value)) {
+    return true;
+  }
+  return coerceSecretRef(value) !== null;
+}
+
+/** Format a ref label without revealing a resolved secret value. */
+function formatSecretRefLabel(ref: SecretRef): string {
+  return `${ref.source}:${ref.provider}:${ref.id}`;
+}
+
+/** Build the unresolved-ref error used when callers bypass gateway secret resolution. */
+function createUnresolvedSecretInputError(params: { path: string; ref: SecretRef }): Error {
+  return new Error(
+    `${params.path}: unresolved SecretRef "${formatSecretRefLabel(params.ref)}". Resolve this command against an active gateway runtime snapshot before reading it.`,
+  );
+}
+
+/** Return a canonical SecretRef when the input is a supported reference shape. */
+export function resolveSecretInputRef(value: unknown): SecretRef | null {
+  return coerceSecretRef(value);
+}
+
+/** Normalize literal secrets, or throw for refs that still require gateway resolution. */
+export function normalizeResolvedSecretInputString(params: {
+  value: unknown;
+  path: string;
+}): string | undefined {
+  const normalized = normalizeSecretInputString(params.value);
+  if (normalized) {
+    return normalized;
+  }
+  const ref = resolveSecretInputRef(params.value);
+  if (!ref) {
+    return undefined;
+  }
+  throw createUnresolvedSecretInputError({ path: params.path, ref });
+}
+
+/** Normalize env-provided secret values before use. */
+export function normalizeEnvSecretInputString(value: unknown): string | undefined {
+  return normalizeSecretInputString(value);
+}

packages/memory-host-sdk/src/host/secret-input.ts

@@ -0,0 +1,32 @@
+// Memory Host SDK module implements secret input behavior.
+import {
+  hasConfiguredSecretInput,
+  normalizeEnvSecretInputString,
+  normalizeResolvedSecretInputString,
+  resolveSecretInputRef,
+} from "./secret-input-utils.js";
+
+// Memory-specific facade for resolving provider secret input from config.
+
+/** Return true when a configured memory secret contains a literal value or reference. */
+export function hasConfiguredMemorySecretInput(value: unknown): boolean {
+  return hasConfiguredSecretInput(value);
+}
+
+/** Resolve memory secret input, reading env refs directly when available. */
+export function resolveMemorySecretInputString(params: {
+  value: unknown;
+  path: string;
+}): string | undefined {
+  const ref = resolveSecretInputRef(params.value);
+  if (ref?.source === "env") {
+    const envValue = normalizeEnvSecretInputString(process.env[ref.id]);
+    if (envValue) {
+      return envValue;
+    }
+  }
+  return normalizeResolvedSecretInputString({
+    value: params.value,
+    path: params.path,
+  });
+}

packages/memory-host-sdk/src/host/session-files-yield.test.ts

@@ -0,0 +1,54 @@
+// Memory Host SDK tests cover session files yield behavior.
+import { afterEach, describe, expect, it, vi } from "vitest";
+
+const { fileState } = vi.hoisted(() => ({
+  fileState: { raw: "" },
+}));
+
+vi.mock("./fs-utils.js", () => ({
+  readRegularFile: vi.fn(async () => ({
+    buffer: Buffer.from(fileState.raw, "utf-8"),
+  })),
+  statRegularFile: vi.fn(async () => ({
+    missing: false,
+    stat: {
+      mtimeMs: 1,
+      size: Buffer.byteLength(fileState.raw, "utf-8"),
+    },
+  })),
+}));
+
+import { buildSessionEntry } from "./session-files.js";
+
+describe("buildSessionEntry responsiveness", () => {
+  afterEach(() => {
+    fileState.raw = "";
+    vi.clearAllMocks();
+  });
+
+  it("yields while parsing a single large transcript", async () => {
+    fileState.raw = Array.from({ length: 25 }, (_value, index) =>
+      JSON.stringify({
+        type: "message",
+        message: { role: "user", content: `message ${index}` },
+      }),
+    ).join("\n");
+    let immediateRan = false;
+    const immediate = new Promise<void>((resolve) => {
+      setImmediate(() => {
+        immediateRan = true;
+        resolve();
+      });
+    });
+
+    const entry = await buildSessionEntry("/tmp/session.jsonl", {
+      generatedByCronRun: false,
+      generatedByDreamingNarrative: false,
+      parseYieldEveryLines: 10,
+    });
+
+    expect(entry?.lineMap).toHaveLength(25);
+    expect(immediateRan).toBe(true);
+    await immediate;
+  });
+});

packages/memory-host-sdk/src/host/session-files.test.ts

@@ -0,0 +1,915 @@
+// Memory Host SDK tests cover session files behavior.
+import fsSync from "node:fs";
+import os from "node:os";
+import path from "node:path";
+import { afterAll, afterEach, beforeAll, beforeEach, describe, expect, it } from "vitest";
+import { clearConfigCache, clearRuntimeConfigSnapshot } from "./openclaw-runtime-session.js";
+import {
+  buildSessionEntry,
+  listSessionFilesForAgent,
+  listSessionTranscriptCorpusEntriesForAgent,
+  loadSessionTranscriptClassificationForAgent,
+  parseCanonicalSessionSyncTargetFromPath,
+  resolveSessionIdentityForTranscriptFile,
+  resolveSessionFileForSyncTarget,
+  sessionPathForFile,
+  type SessionFileEntry,
+} from "./session-files.js";
+
+function captureStateDirEnv() {
+  const stateDir = process.env.OPENCLAW_STATE_DIR;
+  const configPath = process.env.OPENCLAW_CONFIG_PATH;
+  return {
+    restore() {
+      if (stateDir === undefined) {
+        Reflect.deleteProperty(process.env, "OPENCLAW_STATE_DIR");
+      } else {
+        Reflect.set(process.env, "OPENCLAW_STATE_DIR", stateDir);
+      }
+      if (configPath === undefined) {
+        Reflect.deleteProperty(process.env, "OPENCLAW_CONFIG_PATH");
+      } else {
+        Reflect.set(process.env, "OPENCLAW_CONFIG_PATH", configPath);
+      }
+    },
+  };
+}
+
+let fixtureRoot: string;
+let tmpDir: string;
+let envSnapshot: ReturnType<typeof captureStateDirEnv> | undefined;
+let fixtureId = 0;
+
+beforeAll(() => {
+  fixtureRoot = fsSync.mkdtempSync(path.join(os.tmpdir(), "session-entry-test-"));
+});
+
+afterAll(() => {
+  fsSync.rmSync(fixtureRoot, { recursive: true, force: true });
+});
+
+beforeEach(() => {
+  tmpDir = path.join(fixtureRoot, `case-${fixtureId++}`);
+  fsSync.mkdirSync(tmpDir, { recursive: true });
+  envSnapshot = captureStateDirEnv();
+  Reflect.set(process.env, "OPENCLAW_STATE_DIR", tmpDir);
+  clearRuntimeConfigSnapshot();
+  clearConfigCache();
+});
+
+afterEach(() => {
+  envSnapshot?.restore();
+  envSnapshot = undefined;
+  clearRuntimeConfigSnapshot();
+  clearConfigCache();
+});
+
+function requireSessionEntry(entry: SessionFileEntry | null): SessionFileEntry {
+  if (!entry) {
+    throw new Error("expected session entry");
+  }
+  return entry;
+}
+
+describe("listSessionFilesForAgent", () => {
+  it("includes reset and deleted transcripts in session file listing", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    fsSync.mkdirSync(path.join(sessionsDir, "archive"), { recursive: true });
+
+    const included = [
+      "active.jsonl",
+      "active.jsonl.reset.2026-02-16T22-26-33.000Z",
+      "active.jsonl.deleted.2026-02-16T22-27-33.000Z",
+    ];
+    const excluded = ["active.jsonl.bak.2026-02-16T22-28-33.000Z", "sessions.json", "notes.md"];
+    excluded.push("active.checkpoint.11111111-1111-4111-8111-111111111111.jsonl");
+
+    for (const fileName of [...included, ...excluded]) {
+      fsSync.writeFileSync(path.join(sessionsDir, fileName), "");
+    }
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "archive", "nested.jsonl.deleted.2026-02-16T22-29-33.000Z"),
+      "",
+    );
+
+    const files = await listSessionFilesForAgent("main");
+
+    expect(files.map((filePath) => path.basename(filePath)).toSorted()).toEqual(
+      included.toSorted(),
+    );
+  });
+});
+
+describe("listSessionTranscriptCorpusEntriesForAgent", () => {
+  it("lists active session entries with accessor-backed identity and classification", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(path.join(sessionsDir, "narrative.jsonl"), "");
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:dreaming-narrative-run-1": {
+          sessionFile: "narrative.jsonl",
+          sessionId: "narrative",
+        },
+      }),
+    );
+
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual([
+      {
+        agentId: "main",
+        artifactKind: "active-session",
+        generatedByDreamingNarrative: true,
+        sessionFile: path.join(sessionsDir, "narrative.jsonl"),
+        sessionId: "narrative",
+        sessionKey: "agent:main:dreaming-narrative-run-1",
+      },
+    ]);
+  });
+
+  it("keeps archive artifacts in the corpus and inherits active session classification", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    const activePath = path.join(sessionsDir, "cron-run.jsonl");
+    const archivePath = path.join(sessionsDir, "cron-run.jsonl.deleted.2026-02-16T22-27-33.000Z");
+    fsSync.writeFileSync(activePath, "");
+    fsSync.writeFileSync(archivePath, "");
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:cron:job-1:run:run-1": {
+          sessionFile: "cron-run.jsonl",
+          sessionId: "cron-run",
+        },
+      }),
+    );
+
+    const classification = loadSessionTranscriptClassificationForAgent("main");
+
+    expect(classification.cronRunTranscriptPaths).toEqual(
+      new Set([activePath, archivePath].map((filePath) => path.resolve(filePath))),
+    );
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toContainEqual({
+      agentId: "main",
+      artifactKind: "archive-artifact",
+      generatedByCronRun: true,
+      sessionFile: archivePath,
+      sessionId: "cron-run",
+    });
+  });
+
+  it("classifies active entries through cron parentage chains", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    const cronPath = path.join(sessionsDir, "cron-run.jsonl");
+    const spawnedChildPath = path.join(sessionsDir, "spawned-child.jsonl");
+    const keyedChildPath = path.join(sessionsDir, "keyed-child.jsonl");
+    const orphanChildPath = path.join(sessionsDir, "orphan-child.jsonl");
+    const normalPath = path.join(sessionsDir, "normal-child.jsonl");
+    for (const filePath of [
+      cronPath,
+      spawnedChildPath,
+      keyedChildPath,
+      orphanChildPath,
+      normalPath,
+    ]) {
+      fsSync.writeFileSync(filePath, "");
+    }
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:cron:job-1:run:run-1": {
+          sessionFile: "cron-run.jsonl",
+          sessionId: "cron-run",
+        },
+        "agent:main:subagent:spawned-child": {
+          sessionFile: "spawned-child.jsonl",
+          sessionId: "spawned-child",
+          spawnedBy: "agent:main:cron:job-1:run:run-1",
+        },
+        "agent:main:subagent:keyed-child": {
+          parentSessionKey: "agent:main:subagent:spawned-child",
+          sessionFile: "keyed-child.jsonl",
+          sessionId: "keyed-child",
+        },
+        "agent:main:subagent:orphan-child": {
+          sessionFile: "orphan-child.jsonl",
+          sessionId: "orphan-child",
+          spawnedBy: "agent:main:cron:job-1:run:missing",
+        },
+        "agent:main:subagent:normal-child": {
+          sessionFile: "normal-child.jsonl",
+          sessionId: "normal-child",
+          spawnedBy: "agent:main:chat:manual",
+        },
+      }),
+    );
+
+    const classification = loadSessionTranscriptClassificationForAgent("main");
+
+    expect(classification.cronRunTranscriptPaths).toEqual(
+      new Set(
+        [cronPath, spawnedChildPath, keyedChildPath, orphanChildPath].map((filePath) =>
+          path.resolve(filePath),
+        ),
+      ),
+    );
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual(
+      expect.arrayContaining([
+        expect.objectContaining({
+          generatedByCronRun: true,
+          sessionFile: spawnedChildPath,
+          sessionKey: "agent:main:subagent:spawned-child",
+        }),
+        expect.objectContaining({
+          generatedByCronRun: true,
+          sessionFile: keyedChildPath,
+          sessionKey: "agent:main:subagent:keyed-child",
+        }),
+        expect.objectContaining({
+          generatedByCronRun: true,
+          sessionFile: orphanChildPath,
+          sessionKey: "agent:main:subagent:orphan-child",
+        }),
+        expect.objectContaining({
+          sessionFile: normalPath,
+          sessionKey: "agent:main:subagent:normal-child",
+        }),
+      ]),
+    );
+    const entries = await listSessionTranscriptCorpusEntriesForAgent("main");
+    expect(entries.find((entry) => entry.sessionFile === normalPath)?.generatedByCronRun).toBe(
+      undefined,
+    );
+  });
+
+  it("keeps archive classification when the active transcript is missing", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    const archivePath = path.join(sessionsDir, "cron-run.jsonl.reset.2026-02-16T22-26-33.000Z");
+    fsSync.writeFileSync(archivePath, "");
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:cron:job-1:run:run-1": {
+          sessionFile: "cron-run.jsonl",
+          sessionId: "cron-run",
+        },
+      }),
+    );
+
+    const expectedArchivePath = archivePath;
+    const classification = loadSessionTranscriptClassificationForAgent("main");
+
+    expect(classification.cronRunTranscriptPaths).toEqual(new Set([expectedArchivePath]));
+    await expect(listSessionFilesForAgent("main")).resolves.toEqual([expectedArchivePath]);
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual([
+      {
+        agentId: "main",
+        artifactKind: "archive-artifact",
+        generatedByCronRun: true,
+        sessionFile: expectedArchivePath,
+        sessionId: "cron-run",
+      },
+    ]);
+  });
+
+  it("omits active session entries whose transcript files are missing", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:chat:missing": {
+          sessionFile: "missing.jsonl",
+          sessionId: "missing",
+        },
+      }),
+    );
+
+    await expect(listSessionFilesForAgent("main")).resolves.toEqual([]);
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual([]);
+  });
+
+  it("omits active session entries whose transcript path is a symlink", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    const targetPath = path.join(tmpDir, "external.jsonl");
+    const symlinkPath = path.join(sessionsDir, "linked.jsonl");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(targetPath, "");
+    fsSync.symlinkSync(targetPath, symlinkPath);
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:chat:linked": {
+          sessionFile: "linked.jsonl",
+          sessionId: "linked",
+        },
+      }),
+    );
+
+    await expect(listSessionFilesForAgent("main")).resolves.toEqual([]);
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual([]);
+  });
+
+  it("rejects session ids that would escape the sessions directory", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(path.join(tmpDir, "secret.jsonl"), "");
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:chat:escape": {
+          sessionId: "../secret",
+        },
+      }),
+    );
+
+    await expect(listSessionFilesForAgent("main")).resolves.toEqual([]);
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual([]);
+  });
+
+  it("does not classify a fallback transcript when explicit sessionFile is invalid", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    const sessionFile = path.join(sessionsDir, "active.jsonl");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(sessionFile, "");
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:cron:job-1:run:run-1": {
+          sessionFile: "../old.jsonl",
+          sessionId: "active",
+        },
+      }),
+    );
+
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual([
+      {
+        agentId: "main",
+        artifactKind: "orphan-file-artifact",
+        sessionFile,
+        sessionId: "active",
+      },
+    ]);
+  });
+
+  it("rejects relative sessionFile values that escape through nested segments", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    const secretPath = path.join(tmpDir, "agents", "main", "secret.jsonl");
+    fsSync.mkdirSync(path.join(sessionsDir, "sub"), { recursive: true });
+    fsSync.writeFileSync(secretPath, "");
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:chat:escape-file": {
+          sessionFile: "sub/../../secret.jsonl",
+          sessionId: "secret",
+        },
+      }),
+    );
+
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual([]);
+  });
+
+  it("rejects absolute transcript paths owned by another agent", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    const otherSessionsDir = path.join(tmpDir, "agents", "ops", "sessions");
+    const otherSessionFile = path.join(otherSessionsDir, "private.jsonl");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.mkdirSync(otherSessionsDir, { recursive: true });
+    fsSync.writeFileSync(otherSessionFile, "");
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:chat:cross-agent": {
+          sessionFile: otherSessionFile,
+          sessionId: "private",
+        },
+      }),
+    );
+
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual([]);
+  });
+
+  it("falls back to transcript filename identity when an active row lacks sessionId", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    const sessionFile = path.join(sessionsDir, "active-thread-456.jsonl");
+    fsSync.writeFileSync(sessionFile, "");
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:chat:thread-456": {
+          sessionFile: "active-thread-456.jsonl",
+        },
+      }),
+    );
+
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual([
+      {
+        agentId: "main",
+        artifactKind: "active-session",
+        sessionFile,
+        sessionId: "active-thread-456",
+        sessionKey: "agent:main:chat:thread-456",
+      },
+    ]);
+  });
+
+  it("lists only the requested agent's active transcripts from a shared custom store", async () => {
+    const sessionsDir = path.join(tmpDir, "custom-sessions");
+    const sessionFile = path.join(sessionsDir, "custom-thread.jsonl");
+    const otherSessionFile = path.join(sessionsDir, "ops-thread.jsonl");
+    const storePath = path.join(sessionsDir, "sessions.json");
+    const configPath = path.join(tmpDir, "openclaw.json");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(sessionFile, "");
+    fsSync.writeFileSync(otherSessionFile, "");
+    fsSync.writeFileSync(
+      storePath,
+      JSON.stringify({
+        "agent:main:chat:custom": {
+          sessionFile: "custom-thread.jsonl",
+          sessionId: "custom-thread",
+        },
+        "agent:ops:chat:custom": {
+          sessionFile: "ops-thread.jsonl",
+          sessionId: "ops-thread",
+        },
+      }),
+    );
+    fsSync.writeFileSync(configPath, JSON.stringify({ session: { store: storePath } }));
+    Reflect.set(process.env, "OPENCLAW_CONFIG_PATH", configPath);
+    clearRuntimeConfigSnapshot();
+    clearConfigCache();
+
+    await expect(listSessionFilesForAgent("main")).resolves.toEqual([sessionFile]);
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual([
+      {
+        agentId: "main",
+        artifactKind: "active-session",
+        sessionFile,
+        sessionId: "custom-thread",
+        sessionKey: "agent:main:chat:custom",
+      },
+    ]);
+    await expect(listSessionFilesForAgent("ops")).resolves.toEqual([otherSessionFile]);
+  });
+
+  it("keeps unowned archives from an agent-owned fixed session store", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    const archivePath = path.join(sessionsDir, "retained.jsonl.deleted.2026-02-16T22-27-33.000Z");
+    const configPath = path.join(tmpDir, "openclaw.json");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(archivePath, "");
+    fsSync.writeFileSync(path.join(sessionsDir, "sessions.json"), "{}");
+    fsSync.writeFileSync(
+      configPath,
+      JSON.stringify({ session: { store: path.join(sessionsDir, "sessions.json") } }),
+    );
+    Reflect.set(process.env, "OPENCLAW_CONFIG_PATH", configPath);
+    clearRuntimeConfigSnapshot();
+    clearConfigCache();
+
+    await expect(listSessionFilesForAgent("main")).resolves.toEqual([archivePath]);
+    await expect(listSessionTranscriptCorpusEntriesForAgent("main")).resolves.toEqual([
+      {
+        agentId: "main",
+        artifactKind: "archive-artifact",
+        sessionFile: archivePath,
+        sessionId: "retained",
+      },
+    ]);
+  });
+
+  it("resolves absolute transcript paths from a fixed custom store", async () => {
+    const storeDir = path.join(tmpDir, "custom-sessions");
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    const sessionFile = path.join(sessionsDir, "absolute-thread.jsonl");
+    const archivePath = path.join(
+      sessionsDir,
+      "absolute-thread.jsonl.deleted.2026-02-16T22-27-33.000Z",
+    );
+    const storePath = path.join(storeDir, "sessions.json");
+    const configPath = path.join(tmpDir, "openclaw.json");
+    fsSync.mkdirSync(storeDir, { recursive: true });
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(sessionFile, "");
+    fsSync.writeFileSync(archivePath, "");
+    fsSync.writeFileSync(
+      storePath,
+      JSON.stringify({
+        "agent:main:chat:absolute": {
+          sessionFile,
+          sessionId: "absolute-thread",
+        },
+      }),
+    );
+    fsSync.writeFileSync(configPath, JSON.stringify({ session: { store: storePath } }));
+    Reflect.set(process.env, "OPENCLAW_CONFIG_PATH", configPath);
+    clearRuntimeConfigSnapshot();
+    clearConfigCache();
+
+    await expect(listSessionFilesForAgent("main")).resolves.toEqual([sessionFile, archivePath]);
+  });
+
+  it("keeps legacy session keys in non-main per-agent stores", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "ops", "sessions");
+    const sessionFile = path.join(sessionsDir, "legacy-thread.jsonl");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(sessionFile, "");
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "slack:workspace:thread": {
+          sessionFile: "legacy-thread.jsonl",
+          sessionId: "legacy-thread",
+        },
+      }),
+    );
+
+    await expect(listSessionFilesForAgent("ops")).resolves.toEqual([sessionFile]);
+    await expect(listSessionFilesForAgent("main")).resolves.toEqual([]);
+  });
+
+  it("keeps legacy main aliases in a renamed default agent store", async () => {
+    const sessionsDir = path.join(tmpDir, "agents", "ops", "sessions");
+    const sessionFile = path.join(sessionsDir, "legacy-main.jsonl");
+    const configPath = path.join(tmpDir, "openclaw.json");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(sessionFile, "");
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:main": {
+          sessionFile: "legacy-main.jsonl",
+          sessionId: "legacy-main",
+        },
+      }),
+    );
+    fsSync.writeFileSync(
+      configPath,
+      JSON.stringify({ agents: { list: [{ id: "ops", default: true }] } }),
+    );
+    Reflect.set(process.env, "OPENCLAW_CONFIG_PATH", configPath);
+    clearRuntimeConfigSnapshot();
+    clearConfigCache();
+
+    await expect(listSessionFilesForAgent("ops")).resolves.toEqual([sessionFile]);
+  });
+});
+
+describe("sessionPathForFile", () => {
+  it("includes the owning agent id when the transcript lives under an agent sessions dir", () => {
+    const absPath = path.join(
+      tmpDir,
+      "agents",
+      "main",
+      "sessions",
+      "deleted-session.jsonl.deleted.2026-02-16T22-27-33.000Z",
+    );
+
+    expect(sessionPathForFile(absPath)).toBe(
+      "sessions/main/deleted-session.jsonl.deleted.2026-02-16T22-27-33.000Z",
+    );
+  });
+
+  it("keeps the legacy basename-only path when the agent owner cannot be derived", () => {
+    expect(sessionPathForFile(path.join(tmpDir, "loose-session.jsonl"))).toBe(
+      "sessions/loose-session.jsonl",
+    );
+  });
+});
+
+describe("memory session sync targets", () => {
+  it("parses deprecated canonical OpenClaw transcript paths into sync identity", () => {
+    const sessionFile = path.join(tmpDir, "agents", "main", "sessions", "active.jsonl");
+    fsSync.mkdirSync(path.dirname(sessionFile), { recursive: true });
+
+    expect(parseCanonicalSessionSyncTargetFromPath(sessionFile)).toEqual({
+      agentId: "main",
+      sessionId: "active",
+    });
+  });
+
+  it("rejects arbitrary deprecated transcript path hints", () => {
+    expect(parseCanonicalSessionSyncTargetFromPath(path.join(tmpDir, "active.jsonl"))).toBeNull();
+    expect(
+      parseCanonicalSessionSyncTargetFromPath(
+        path.join(tmpDir, "agents", "main", "sessions", "active.trajectory.jsonl"),
+      ),
+    ).toBeNull();
+  });
+
+  it("resolves identity sync targets to the current file-backed transcript", () => {
+    expect(resolveSessionFileForSyncTarget({ sessionId: "active" }, "main")).toEqual({
+      agentId: "main",
+      sessionId: "active",
+      sessionFile: path.join(tmpDir, "agents", "main", "sessions", "active.jsonl"),
+    });
+  });
+
+  it("normalizes agent ids before resolving identity sync targets", () => {
+    expect(resolveSessionFileForSyncTarget({ agentId: "MAIN", sessionId: "active" })).toEqual({
+      agentId: "main",
+      sessionId: "active",
+      sessionFile: path.join(tmpDir, "agents", "main", "sessions", "active.jsonl"),
+    });
+  });
+
+  it("rejects identity sync targets that would escape the sessions directory", () => {
+    expect(resolveSessionFileForSyncTarget({ sessionId: "../outside" }, "main")).toBeNull();
+  });
+
+  it("rejects identity sync targets that normalize to another transcript", () => {
+    expect(resolveSessionFileForSyncTarget({ sessionId: "foo/../active" }, "main")).toBeNull();
+  });
+
+  it("resolves identity sync targets through persisted session keys", () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:chat:thread-456": {
+          sessionFile: "active-thread-456.jsonl",
+          sessionId: "active",
+        },
+      }),
+    );
+
+    expect(
+      resolveSessionFileForSyncTarget({
+        agentId: "main",
+        sessionId: "active",
+        sessionKey: "agent:main:chat:thread-456",
+      }),
+    ).toEqual({
+      agentId: "main",
+      sessionId: "active",
+      sessionFile: path.join(sessionsDir, "active-thread-456.jsonl"),
+    });
+  });
+
+  it("resolves identity sync targets through persisted session ids", () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:chat:thread-456": {
+          sessionFile: "active-thread-456.jsonl",
+          sessionId: "active",
+        },
+      }),
+    );
+
+    expect(resolveSessionFileForSyncTarget({ agentId: "main", sessionId: "active" })).toEqual({
+      agentId: "main",
+      sessionId: "active",
+      sessionFile: path.join(sessionsDir, "active-thread-456.jsonl"),
+    });
+  });
+
+  it("resolves transcript file identities through persisted session keys", () => {
+    const sessionsDir = path.join(tmpDir, "agents", "main", "sessions");
+    fsSync.mkdirSync(sessionsDir, { recursive: true });
+    const sessionFile = path.join(sessionsDir, "active-thread-456.jsonl");
+    fsSync.writeFileSync(
+      path.join(sessionsDir, "sessions.json"),
+      JSON.stringify({
+        "agent:main:chat:thread-456": {
+          sessionFile: "active-thread-456.jsonl",
+          sessionId: "active",
+        },
+      }),
+    );
+
+    expect(resolveSessionIdentityForTranscriptFile(sessionFile)).toEqual({
+      agentId: "main",
+      sessionId: "active",
+      sessionKey: "agent:main:chat:thread-456",
+    });
+  });
+});
+
+describe("buildSessionEntry", () => {
+  it("returns lineMap tracking original JSONL line numbers", async () => {
+    // Simulate a real session JSONL file with metadata records interspersed
+    // Lines 1-3: non-message metadata records
+    // Line 4: user message
+    // Line 5: metadata
+    // Line 6: assistant message
+    // Line 7: user message
+    const jsonlLines = [
+      JSON.stringify({ type: "custom", customType: "model-snapshot", data: {} }),
+      JSON.stringify({ type: "custom", customType: "openclaw.cache-ttl", data: {} }),
+      JSON.stringify({ type: "session-meta", agentId: "test" }),
+      JSON.stringify({ type: "message", message: { role: "user", content: "Hello world" } }),
+      JSON.stringify({ type: "custom", customType: "tool-result", data: {} }),
+      JSON.stringify({
+        type: "message",
+        message: { role: "assistant", content: "Hi there, how can I help?" },
+      }),
+      JSON.stringify({ type: "message", message: { role: "user", content: "Tell me a joke" } }),
+    ];
+    const filePath = path.join(tmpDir, "session.jsonl");
+    fsSync.writeFileSync(filePath, jsonlLines.join("\n"));
+
+    const entry = requireSessionEntry(await buildSessionEntry(filePath));
+    expect(entry.content).toBe(
+      "User: Hello world\nAssistant: Hi there, how can I help?\nUser: Tell me a joke",
+    );
+
+    // lineMap should map each content line to its original JSONL line (1-indexed)
+    // Content line 0 → JSONL line 4 (the first user message)
+    // Content line 1 → JSONL line 6 (the assistant message)
+    // Content line 2 → JSONL line 7 (the second user message)
+    expect(entry.lineMap).toStrictEqual([4, 6, 7]);
+  });
+
+  it("returns empty lineMap when no messages are found", async () => {
+    const jsonlLines = [
+      JSON.stringify({ type: "custom", customType: "model-snapshot", data: {} }),
+      JSON.stringify({ type: "session-meta", agentId: "test" }),
+    ];
+    const filePath = path.join(tmpDir, "empty-session.jsonl");
+    fsSync.writeFileSync(filePath, jsonlLines.join("\n"));
+
+    const entry = requireSessionEntry(await buildSessionEntry(filePath));
+    expect(entry.content).toBe("");
+    expect(entry.lineMap).toStrictEqual([]);
+  });
+
+  it("indexes usage-counted reset/deleted archives but still skips bak and checkpoint artifacts", async () => {
+    const resetPath = path.join(tmpDir, "ordinary.jsonl.reset.2026-02-16T22-26-33.000Z");
+    const deletedPath = path.join(tmpDir, "ordinary.jsonl.deleted.2026-02-16T22-27-33.000Z");
+    const bakPath = path.join(tmpDir, "ordinary.jsonl.bak.2026-02-16T22-28-33.000Z");
+    const checkpointPath = path.join(
+      tmpDir,
+      "ordinary.checkpoint.11111111-1111-4111-8111-111111111111.jsonl",
+    );
+    const content = JSON.stringify({
+      type: "message",
+      message: { role: "user", content: "Archived hello" },
+    });
+    fsSync.writeFileSync(resetPath, content);
+    fsSync.writeFileSync(deletedPath, content);
+    fsSync.writeFileSync(bakPath, content);
+    fsSync.writeFileSync(checkpointPath, content);
+
+    const resetEntry = requireSessionEntry(await buildSessionEntry(resetPath));
+    const deletedEntry = requireSessionEntry(await buildSessionEntry(deletedPath));
+    const bakEntry = requireSessionEntry(await buildSessionEntry(bakPath));
+    const checkpointEntry = requireSessionEntry(await buildSessionEntry(checkpointPath));
+
+    // Usage-counted archives (reset, deleted) must surface real content so
+    // post-reset memory_search can recover prior session history.
+    expect(resetEntry.content).toBe("User: Archived hello");
+    expect(resetEntry.lineMap).toStrictEqual([1]);
+    expect(deletedEntry.content).toBe("User: Archived hello");
+    expect(deletedEntry.lineMap).toStrictEqual([1]);
+
+    // .bak and compaction checkpoints remain opaque pre-archive / snapshot
+    // artifacts and stay empty so they do not get double-indexed.
+    expect(bakEntry.content).toBe("");
+    expect(bakEntry.lineMap).toStrictEqual([]);
+    expect(checkpointEntry.content).toBe("");
+    expect(checkpointEntry.lineMap).toStrictEqual([]);
+  });
+
+  it("keeps cron-run deleted archives opaque when the live session store entry is gone", async () => {
+    const archivePath = path.join(tmpDir, "cron-run.jsonl.deleted.2026-02-16T22-27-33.000Z");
+    const jsonlLines = [
+      JSON.stringify({
+        type: "message",
+        message: {
+          role: "user",
+          content: "[cron:job-1 Codex Sessions Sync] Run internal sync.",
+        },
+      }),
+      JSON.stringify({
+        type: "message",
+        message: { role: "assistant", content: "Internal cron output that must stay out." },
+      }),
+    ];
+    fsSync.writeFileSync(archivePath, jsonlLines.join("\n"));
+
+    const entry = requireSessionEntry(await buildSessionEntry(archivePath));
+
+    expect(entry.content).toBe("");
+    expect(entry.lineMap).toStrictEqual([]);
+    expect(entry.generatedByCronRun).toBe(true);
+  });
+
+  it("keeps cron-run reset archives opaque when session metadata preserves the cron key", async () => {
+    const archivePath = path.join(tmpDir, "cron-run.jsonl.reset.2026-02-16T22-26-33.000Z");
+    const jsonlLines = [
+      JSON.stringify({
+        type: "session-meta",
+        data: { sessionKey: "agent:main:cron:job-1:run:run-1" },
+      }),
+      JSON.stringify({
+        type: "message",
+        message: { role: "assistant", content: "Internal cron output that must stay out." },
+      }),
+    ];
+    fsSync.writeFileSync(archivePath, jsonlLines.join("\n"));
+
+    const entry = requireSessionEntry(await buildSessionEntry(archivePath));
+
+    expect(entry.content).toBe("");
+    expect(entry.lineMap).toStrictEqual([]);
+    expect(entry.generatedByCronRun).toBe(true);
+  });
+
+  it("skips blank lines and invalid JSON without breaking lineMap", async () => {
+    const jsonlLines = [
+      "",
+      "not valid json",
+      JSON.stringify({ type: "message", message: { role: "user", content: "First" } }),
+      "",
+      JSON.stringify({ type: "message", message: { role: "assistant", content: "Second" } }),
+    ];
+    const filePath = path.join(tmpDir, "gaps.jsonl");
+    fsSync.writeFileSync(filePath, jsonlLines.join("\n"));
+
+    const entry = requireSessionEntry(await buildSessionEntry(filePath));
+    expect(entry.lineMap).toStrictEqual([3, 5]);
+  });
+
+  it("strips inbound metadata when a user envelope is split across text blocks", async () => {
+    const jsonlLines = [
+      JSON.stringify({
+        type: "message",
+        message: {
+          role: "user",
+          content: [
+            { type: "text", text: "Conversation info (untrusted metadata):" },
+            { type: "text", text: "```json" },
+            { type: "text", text: '{"message_id":"msg-100","chat_id":"-100123"}' },
+            { type: "text", text: "```" },
+            { type: "text", text: "" },
+            { type: "text", text: "Sender (untrusted metadata):" },
+            { type: "text", text: "```json" },
+            { type: "text", text: '{"label":"Chris","id":"42"}' },
+            { type: "text", text: "```" },
+            { type: "text", text: "" },
+            { type: "text", text: "Actual user text" },
+          ],
+        },
+      }),
+    ];
+    const filePath = path.join(tmpDir, "enveloped-session-array.jsonl");
+    fsSync.writeFileSync(filePath, jsonlLines.join("\n"));
+
+    const entry = requireSessionEntry(await buildSessionEntry(filePath));
+    expect(entry.content).toBe("User: Actual user text");
+  });
+
+  it("skips inter-session user messages", async () => {
+    const jsonlLines = [
+      JSON.stringify({
+        type: "message",
+        message: {
+          role: "user",
+          content: "A background task completed. Internal relay text.",
+          provenance: { kind: "inter_session", sourceTool: "subagent_announce" },
+        },
+      }),
+      JSON.stringify({
+        type: "message",
+        message: { role: "assistant", content: "User-facing summary." },
+      }),
+      JSON.stringify({
+        type: "message",
+        message: { role: "user", content: "Actual user follow-up." },
+      }),
+    ];
+    const filePath = path.join(tmpDir, "inter-session-session.jsonl");
+    fsSync.writeFileSync(filePath, jsonlLines.join("\n"));
+
+    const entry = requireSessionEntry(await buildSessionEntry(filePath));
+    expect(entry.content).toBe("Assistant: User-facing summary.\nUser: Actual user follow-up.");
+    expect(entry.lineMap).toStrictEqual([2, 3]);
+  });
+
+  it("drops Date-invalid numeric message timestamps", async () => {
+    const jsonlLines = [
+      JSON.stringify({
+        type: "message",
+        message: {
+          role: "user",
+          content: "Hello",
+          timestamp: 8_640_000_000_000_001,
+        },
+      }),
+    ];
+    const filePath = path.join(tmpDir, "invalid-timestamp-session.jsonl");
+    fsSync.writeFileSync(filePath, jsonlLines.join("\n"));
+
+    const entry = requireSessionEntry(await buildSessionEntry(filePath));
+    expect(entry.messageTimestampsMs).toStrictEqual([0]);
+  });
+});

packages/memory-host-sdk/src/host/session-files.ts

@@ -0,0 +1,870 @@
+// Memory Host SDK module implements session files behavior.
+import fsSync from "node:fs";
+import path from "node:path";
+import { normalizeAgentId } from "./config-utils.js";
+import { readRegularFile, statRegularFile } from "./fs-utils.js";
+import { hashText } from "./hash.js";
+import { createSubsystemLogger, redactSensitiveText } from "./openclaw-runtime-io.js";
+import {
+  DREAMING_NARRATIVE_RUN_PREFIX,
+  isDreamingNarrativeSessionStoreKey,
+  extractAgentIdFromSessionsDir,
+  HEARTBEAT_PROMPT,
+  HEARTBEAT_TOKEN,
+  hasInterSessionUserProvenance,
+  isCompactionCheckpointTranscriptFileName,
+  isCronRunSessionKey,
+  isExecCompletionEvent,
+  isHeartbeatUserMessage,
+  isSessionArchiveArtifactName,
+  isSilentReplyPayloadText,
+  isUsageCountedSessionTranscriptFileName,
+  parseUsageCountedSessionIdFromFileName,
+  resolveSessionTranscriptsDirForAgent,
+  stripInboundMetadata,
+  stripInternalRuntimeContext,
+} from "./openclaw-runtime-session.js";
+import { retryTransientMemoryRead } from "./read-retry.js";
+import {
+  listSessionTranscriptCorpusEntriesForAgent,
+  listSessionTranscriptCorpusEntriesForAgentSync,
+  type SessionTranscriptCorpusEntry,
+} from "./session-transcript-corpus.js";
+import type { MemorySessionSyncTarget } from "./types.js";
+
+export {
+  listSessionTranscriptCorpusEntriesForAgent,
+  type SessionTranscriptCorpusArtifactKind,
+  type SessionTranscriptCorpusEntry,
+} from "./session-transcript-corpus.js";
+
+// Keep the historical one-line-per-message export shape for normal turns, but
+// wrap pathological long messages so downstream indexers never ingest a single
+// toxic line. Wrapped continuation lines still map back to the same JSONL line.
+// This limit applies to content only; the role label adds up to 11 chars.
+const SESSION_EXPORT_CONTENT_WRAP_CHARS = 800;
+const SESSION_ENTRY_PARSE_YIELD_LINES = 250;
+const MAX_DATE_TIMESTAMP_MS = 8_640_000_000_000_000;
+const DIRECT_CRON_PROMPT_RE = /^\[cron:[^\]]+\]\s*/;
+
+export type SessionFileEntry = {
+  path: string;
+  absPath: string;
+  mtimeMs: number;
+  size: number;
+  hash: string;
+  content: string;
+  /** Maps each content line (0-indexed) to its 1-indexed JSONL source line. */
+  lineMap: number[];
+  /** Maps each content line (0-indexed) to epoch ms; 0 means unknown timestamp. */
+  messageTimestampsMs: number[];
+  /** True when this transcript belongs to an internal dreaming narrative run. */
+  generatedByDreamingNarrative?: boolean;
+  /** True when this transcript belongs to an isolated cron run session. */
+  generatedByCronRun?: boolean;
+};
+
+export type BuildSessionEntryOptions = {
+  /** Optional preclassification from a caller-managed dreaming transcript lookup. */
+  generatedByDreamingNarrative?: boolean;
+  /** Optional preclassification from a caller-managed cron transcript lookup. */
+  generatedByCronRun?: boolean;
+  /** Override for tests or specialized callers that need a tighter parse yield cadence. */
+  parseYieldEveryLines?: number;
+};
+
+export type SessionTranscriptClassification = {
+  dreamingNarrativeTranscriptPaths: ReadonlySet<string>;
+  cronRunTranscriptPaths: ReadonlySet<string>;
+};
+
+export type ResolvedMemorySessionSyncTarget = {
+  agentId: string;
+  sessionFile: string;
+  sessionId: string;
+};
+
+export type ResolvedSessionTranscriptIdentity = {
+  agentId: string;
+  sessionId: string;
+  sessionKey?: string;
+};
+
+type SessionTranscriptStoreEntry = {
+  sessionFile?: unknown;
+  sessionId?: unknown;
+};
+
+function shouldSkipTranscriptFileForDreaming(absPath: string): boolean {
+  const fileName = path.basename(absPath);
+  // Compaction checkpoints are always skipped: they are derived snapshots of an
+  // active session and would double-index the same content.
+  if (isCompactionCheckpointTranscriptFileName(fileName)) {
+    return true;
+  }
+  // Legacy backups and `.jsonl.bak.<iso>` rotations are opaque pre-archive
+  // copies, not a user-facing session artifact; skip them too.
+  if (
+    isSessionArchiveArtifactName(fileName) &&
+    !isUsageCountedSessionTranscriptFileName(fileName)
+  ) {
+    return true;
+  }
+  // Usage-counted archives (`.jsonl.reset.<iso>` / `.jsonl.deleted.<iso>`) are
+  // the rotated-but-retained copies of real sessions and must stay indexed so
+  // `memory_search` can surface hits on post-reset / post-delete history.
+  return false;
+}
+
+function isUsageCountedSessionArchiveTranscriptPath(absPath: string): boolean {
+  const fileName = path.basename(absPath);
+  return (
+    isUsageCountedSessionTranscriptFileName(fileName) &&
+    isSessionArchiveArtifactName(fileName) &&
+    parseUsageCountedSessionIdFromFileName(fileName) !== null
+  );
+}
+
+function isDreamingNarrativeBootstrapRecord(record: unknown): boolean {
+  if (!record || typeof record !== "object" || Array.isArray(record)) {
+    return false;
+  }
+  const candidate = record as {
+    type?: unknown;
+    customType?: unknown;
+    data?: unknown;
+  };
+  if (
+    candidate.type !== "custom" ||
+    candidate.customType !== "openclaw:bootstrap-context:full" ||
+    !candidate.data ||
+    typeof candidate.data !== "object" ||
+    Array.isArray(candidate.data)
+  ) {
+    return false;
+  }
+  const runId = (candidate.data as { runId?: unknown }).runId;
+  return typeof runId === "string" && runId.startsWith(DREAMING_NARRATIVE_RUN_PREFIX);
+}
+
+function hasDreamingNarrativeRunId(value: unknown): boolean {
+  return typeof value === "string" && value.startsWith(DREAMING_NARRATIVE_RUN_PREFIX);
+}
+
+function isDreamingNarrativeGeneratedRecord(record: unknown): boolean {
+  if (isDreamingNarrativeBootstrapRecord(record)) {
+    return true;
+  }
+  if (!record || typeof record !== "object" || Array.isArray(record)) {
+    return false;
+  }
+  const candidate = record as {
+    runId?: unknown;
+    sessionKey?: unknown;
+    data?: unknown;
+  };
+  if (
+    hasDreamingNarrativeRunId(candidate.runId) ||
+    hasDreamingNarrativeRunId(candidate.sessionKey)
+  ) {
+    return true;
+  }
+  if (!candidate.data || typeof candidate.data !== "object" || Array.isArray(candidate.data)) {
+    return false;
+  }
+  const nested = candidate.data as {
+    runId?: unknown;
+    sessionKey?: unknown;
+  };
+  return hasDreamingNarrativeRunId(nested.runId) || hasDreamingNarrativeRunId(nested.sessionKey);
+}
+
+function hasCronRunSessionKey(value: unknown): boolean {
+  return typeof value === "string" && isCronRunSessionKey(value);
+}
+
+function isCronRunGeneratedRecord(record: unknown): boolean {
+  if (!record || typeof record !== "object" || Array.isArray(record)) {
+    return false;
+  }
+  const candidate = record as {
+    sessionKey?: unknown;
+    data?: unknown;
+  };
+  if (hasCronRunSessionKey(candidate.sessionKey)) {
+    return true;
+  }
+  if (!candidate.data || typeof candidate.data !== "object" || Array.isArray(candidate.data)) {
+    return false;
+  }
+  const nested = candidate.data as {
+    sessionKey?: unknown;
+  };
+  return hasCronRunSessionKey(nested.sessionKey);
+}
+
+function normalizeComparablePath(pathname: string): string {
+  const resolved = path.resolve(pathname);
+  return process.platform === "win32" ? resolved.toLowerCase() : resolved;
+}
+
+export function normalizeSessionTranscriptPathForComparison(pathname: string): string {
+  return normalizeComparablePath(pathname);
+}
+
+function resolveSessionStoreTranscriptPath(
+  sessionsDir: string,
+  entry: { sessionFile?: unknown; sessionId?: unknown } | undefined,
+): string | null {
+  const resolved = resolveSessionStoreTranscriptResolvedPath(sessionsDir, entry);
+  return resolved ? normalizeComparablePath(resolved) : null;
+}
+
+function resolveSessionStoreTranscriptResolvedPath(
+  sessionsDir: string,
+  entry: { sessionFile?: unknown; sessionId?: unknown } | undefined,
+): string | null {
+  if (typeof entry?.sessionFile === "string" && entry.sessionFile.trim().length > 0) {
+    const sessionFile = entry.sessionFile.trim();
+    return path.isAbsolute(sessionFile) ? sessionFile : path.resolve(sessionsDir, sessionFile);
+  }
+  if (typeof entry?.sessionId === "string" && entry.sessionId.trim().length > 0) {
+    return path.join(sessionsDir, `${entry.sessionId.trim()}.jsonl`);
+  }
+  return null;
+}
+
+function isCanonicalSessionsDirForAgent(sessionsDir: string, agentId: string): boolean {
+  return (
+    normalizeComparablePath(sessionsDir) ===
+    normalizeComparablePath(resolveSessionTranscriptsDirForAgent(agentId))
+  );
+}
+
+export function loadSessionTranscriptClassificationForSessionsDir(
+  sessionsDir: string,
+): SessionTranscriptClassification {
+  const agentId = extractAgentIdFromSessionsDir(sessionsDir);
+  if (agentId && isCanonicalSessionsDirForAgent(sessionsDir, agentId)) {
+    return classifySessionTranscriptCorpusEntries(
+      listSessionTranscriptCorpusEntriesForAgentSync(agentId),
+    );
+  }
+  const storePath = path.join(sessionsDir, "sessions.json");
+  const store = readSessionTranscriptClassificationStore(storePath);
+  const dreamingTranscriptPaths = new Set<string>();
+  const cronRunTranscriptPaths = new Set<string>();
+  for (const [sessionKey, entry] of Object.entries(store)) {
+    const transcriptPath = resolveSessionStoreTranscriptPath(sessionsDir, entry);
+    if (!transcriptPath) {
+      continue;
+    }
+    if (isDreamingNarrativeSessionStoreKey(sessionKey)) {
+      dreamingTranscriptPaths.add(transcriptPath);
+    }
+    if (isCronRunSessionKey(sessionKey)) {
+      cronRunTranscriptPaths.add(transcriptPath);
+    }
+  }
+  return {
+    dreamingNarrativeTranscriptPaths: dreamingTranscriptPaths,
+    cronRunTranscriptPaths,
+  };
+}
+
+function readSessionTranscriptClassificationStore(
+  storePath: string,
+): Record<string, SessionTranscriptStoreEntry> {
+  try {
+    const parsed = JSON.parse(fsSync.readFileSync(storePath, "utf-8")) as unknown;
+    if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) {
+      return {};
+    }
+    return parsed as Record<string, SessionTranscriptStoreEntry>;
+  } catch {
+    return {};
+  }
+}
+
+function classifySessionTranscriptCorpusEntries(
+  corpusEntries: readonly SessionTranscriptCorpusEntry[],
+): SessionTranscriptClassification {
+  const dreamingTranscriptPaths = new Set<string>();
+  const cronRunTranscriptPaths = new Set<string>();
+  for (const entry of corpusEntries) {
+    const normalizedPath = normalizeComparablePath(entry.sessionFile);
+    if (entry.generatedByDreamingNarrative) {
+      dreamingTranscriptPaths.add(normalizedPath);
+    }
+    if (entry.generatedByCronRun) {
+      cronRunTranscriptPaths.add(normalizedPath);
+    }
+  }
+  return {
+    dreamingNarrativeTranscriptPaths: dreamingTranscriptPaths,
+    cronRunTranscriptPaths,
+  };
+}
+
+function findSessionTranscriptStoreEntryBySessionId(
+  store: Record<string, SessionTranscriptStoreEntry>,
+  sessionId: string,
+): SessionTranscriptStoreEntry | undefined {
+  return Object.values(store).find((entry) => {
+    return typeof entry.sessionId === "string" && entry.sessionId.trim() === sessionId;
+  });
+}
+
+export function loadDreamingNarrativeTranscriptPathSetForAgent(
+  agentId: string,
+): ReadonlySet<string> {
+  return loadSessionTranscriptClassificationForAgent(agentId).dreamingNarrativeTranscriptPaths;
+}
+
+export function loadSessionTranscriptClassificationForAgent(
+  agentId: string,
+): SessionTranscriptClassification {
+  return classifySessionTranscriptCorpusEntries(
+    listSessionTranscriptCorpusEntriesForAgentSync(agentId),
+  );
+}
+
+function classifySessionTranscriptFromSessionStore(absPath: string): {
+  generatedByDreamingNarrative: boolean;
+  generatedByCronRun: boolean;
+} {
+  const sessionsDir = path.dirname(absPath);
+  const normalizedAbsPath = normalizeComparablePath(absPath);
+  const primarySessionId = parseUsageCountedSessionIdFromFileName(path.basename(absPath));
+  const normalizedPrimaryPath =
+    primarySessionId && isSessionArchiveArtifactName(path.basename(absPath))
+      ? normalizeComparablePath(path.join(sessionsDir, `${primarySessionId}.jsonl`))
+      : null;
+  const classification = loadSessionTranscriptClassificationForSessionsDir(sessionsDir);
+  const hasClassifiedPath = (paths: ReadonlySet<string>) =>
+    paths.has(normalizedAbsPath) ||
+    (normalizedPrimaryPath !== null && paths.has(normalizedPrimaryPath));
+  return {
+    generatedByDreamingNarrative: hasClassifiedPath(
+      classification.dreamingNarrativeTranscriptPaths,
+    ),
+    generatedByCronRun: hasClassifiedPath(classification.cronRunTranscriptPaths),
+  };
+}
+
+export async function listSessionFilesForAgent(agentId: string): Promise<string[]> {
+  return (await listSessionTranscriptCorpusEntriesForAgent(agentId)).map(
+    (entry) => entry.sessionFile,
+  );
+}
+
+function extractAgentIdFromSessionPath(absPath: string): string | null {
+  const parts = path.normalize(path.resolve(absPath)).split(path.sep).filter(Boolean);
+  const sessionsIndex = parts.lastIndexOf("sessions");
+  if (sessionsIndex < 2 || parts[sessionsIndex - 2] !== "agents") {
+    return null;
+  }
+  return parts[sessionsIndex - 1] || null;
+}
+
+export function sessionPathForFile(absPath: string): string {
+  const agentId = extractAgentIdFromSessionPath(absPath);
+  return path
+    .join("sessions", ...(agentId ? [agentId] : []), path.basename(absPath))
+    .replace(/\\/g, "/");
+}
+
+/**
+ * Parses a deprecated path-shaped memory sync hint only when it points at an
+ * OpenClaw-owned usage-counted transcript in the canonical agent sessions dir.
+ */
+export function parseCanonicalSessionSyncTargetFromPath(
+  sessionFile: string,
+): MemorySessionSyncTarget | null {
+  const trimmed = sessionFile.trim();
+  if (!trimmed) {
+    return null;
+  }
+  const resolved = path.resolve(trimmed);
+  const fileName = path.basename(resolved);
+  const sessionId = parseUsageCountedSessionIdFromFileName(fileName);
+  if (!sessionId || !isUsageCountedSessionTranscriptFileName(fileName)) {
+    return null;
+  }
+  const agentId = extractAgentIdFromSessionPath(resolved);
+  if (!agentId) {
+    return null;
+  }
+  const canonicalSessionsDir = normalizeComparablePath(
+    resolveSessionTranscriptsDirForAgent(agentId),
+  );
+  if (normalizeComparablePath(path.dirname(resolved)) !== canonicalSessionsDir) {
+    return null;
+  }
+  return { agentId, sessionId };
+}
+
+/**
+ * Resolves a current transcript path back to the canonical session-store
+ * identity when available, falling back to the usage-counted file identity.
+ */
+export function resolveSessionIdentityForTranscriptFile(
+  sessionFile: string,
+): ResolvedSessionTranscriptIdentity | null {
+  const parsed = parseCanonicalSessionSyncTargetFromPath(sessionFile);
+  if (!parsed?.agentId) {
+    return null;
+  }
+  const sessionsDir = resolveSessionTranscriptsDirForAgent(parsed.agentId);
+  const normalizedSessionFile = normalizeComparablePath(sessionFile);
+  const store = readSessionTranscriptClassificationStore(path.join(sessionsDir, "sessions.json"));
+  for (const [sessionKey, entry] of Object.entries(store)) {
+    const transcriptPath = resolveSessionStoreTranscriptPath(sessionsDir, entry);
+    if (transcriptPath !== normalizedSessionFile) {
+      continue;
+    }
+    const sessionId = typeof entry.sessionId === "string" ? entry.sessionId.trim() : "";
+    if (!sessionId) {
+      continue;
+    }
+    return {
+      agentId: parsed.agentId,
+      sessionId,
+      ...(sessionKey.trim() ? { sessionKey } : {}),
+    };
+  }
+  return {
+    agentId: parsed.agentId,
+    sessionId: parsed.sessionId,
+  };
+}
+
+/**
+ * Resolves a storage-neutral memory sync target to the current file-backed
+ * transcript. The SQLite adapter implements this identity contract without
+ * deriving a path.
+ */
+export function resolveSessionFileForSyncTarget(
+  target: MemorySessionSyncTarget,
+  defaultAgentId?: string,
+): ResolvedMemorySessionSyncTarget | null {
+  const sessionId = target.sessionId.trim();
+  const rawAgentId = (target.agentId ?? defaultAgentId ?? "").trim();
+  if (!rawAgentId || !sessionId) {
+    return null;
+  }
+  const agentId = normalizeAgentId(rawAgentId);
+  const sessionsDir = resolveSessionTranscriptsDirForAgent(agentId);
+  const sessionKey = target.sessionKey?.trim();
+  let store: Record<string, SessionTranscriptStoreEntry> | null = null;
+  if (sessionKey) {
+    store = readSessionTranscriptClassificationStore(path.join(sessionsDir, "sessions.json"));
+    const persistedPath = resolveSessionStoreTranscriptResolvedPath(sessionsDir, store[sessionKey]);
+    const canonicalPath = resolveCanonicalSessionSyncFilePath(agentId, persistedPath);
+    if (canonicalPath) {
+      return {
+        agentId,
+        sessionId,
+        sessionFile: canonicalPath,
+      };
+    }
+  }
+  store ??= readSessionTranscriptClassificationStore(path.join(sessionsDir, "sessions.json"));
+  const persistedPath = resolveSessionStoreTranscriptResolvedPath(
+    sessionsDir,
+    findSessionTranscriptStoreEntryBySessionId(store, sessionId),
+  );
+  const canonicalPath = resolveCanonicalSessionSyncFilePath(agentId, persistedPath);
+  if (canonicalPath) {
+    return {
+      agentId,
+      sessionId,
+      sessionFile: canonicalPath,
+    };
+  }
+  const sessionFile = resolveCanonicalSessionSyncFilePath(
+    agentId,
+    path.join(sessionsDir, `${sessionId}.jsonl`),
+    sessionId,
+  );
+  if (!sessionFile) {
+    return null;
+  }
+  return {
+    agentId,
+    sessionId,
+    sessionFile,
+  };
+}
+
+function resolveCanonicalSessionSyncFilePath(
+  agentId: string,
+  sessionFile?: string | null,
+  expectedSessionId?: string,
+): string | null {
+  if (!sessionFile) {
+    return null;
+  }
+  const resolved = path.resolve(sessionFile);
+  const parsed = parseCanonicalSessionSyncTargetFromPath(resolved);
+  if (parsed?.agentId !== agentId) {
+    return null;
+  }
+  if (expectedSessionId !== undefined && parsed.sessionId !== expectedSessionId) {
+    return null;
+  }
+  return resolved;
+}
+
+async function logSessionFileReadFailure(absPath: string, err: unknown): Promise<void> {
+  createSubsystemLogger("memory").debug(`Failed reading session file ${absPath}: ${String(err)}`);
+}
+
+function normalizeSessionText(value: string): string {
+  return value
+    .replace(/\s*\n+\s*/g, " ")
+    .replace(/\s+/g, " ")
+    .trim();
+}
+
+function collectRawSessionText(content: unknown): string | null {
+  if (typeof content === "string") {
+    return content;
+  }
+  if (!Array.isArray(content)) {
+    return null;
+  }
+  const parts: string[] = [];
+  for (const block of content) {
+    if (!block || typeof block !== "object") {
+      continue;
+    }
+    const record = block as { type?: unknown; text?: unknown };
+    if (record.type === "text" && typeof record.text === "string") {
+      parts.push(record.text);
+    }
+  }
+  return parts.length > 0 ? parts.join("\n") : null;
+}
+
+function isHighSurrogate(code: number): boolean {
+  return code >= 0xd800 && code <= 0xdbff;
+}
+
+function isLowSurrogate(code: number): boolean {
+  return code >= 0xdc00 && code <= 0xdfff;
+}
+
+function splitLongSessionLine(
+  text: string,
+  maxChars: number = SESSION_EXPORT_CONTENT_WRAP_CHARS,
+): string[] {
+  const normalized = text.trim();
+  if (!normalized) {
+    return [];
+  }
+  if (normalized.length <= maxChars) {
+    return [normalized];
+  }
+
+  const segments: string[] = [];
+  let cursor = 0;
+  while (cursor < normalized.length) {
+    const remaining = normalized.length - cursor;
+    if (remaining <= maxChars) {
+      segments.push(normalized.slice(cursor).trim());
+      break;
+    }
+
+    const limit = cursor + maxChars;
+    let splitAt = limit;
+    for (let index = limit; index > cursor; index -= 1) {
+      if (normalized[index] === " ") {
+        splitAt = index;
+        break;
+      }
+    }
+    if (
+      splitAt < normalized.length &&
+      splitAt > cursor &&
+      isHighSurrogate(normalized.charCodeAt(splitAt - 1)) &&
+      isLowSurrogate(normalized.charCodeAt(splitAt))
+    ) {
+      splitAt -= 1;
+    }
+    segments.push(normalized.slice(cursor, splitAt).trim());
+    cursor = splitAt;
+    while (cursor < normalized.length && normalized[cursor] === " ") {
+      cursor += 1;
+    }
+  }
+
+  return segments.filter(Boolean);
+}
+
+function renderSessionExportLines(label: string, text: string): string[] {
+  return splitLongSessionLine(text).map((segment) => `${label}: ${segment}`);
+}
+
+/**
+ * Strip OpenClaw-injected inbound metadata envelopes from a raw text block.
+ *
+ * User-role messages arriving from external channels (Telegram, Discord,
+ * Slack, …) are stored with a multi-line prefix containing Conversation info,
+ * Sender info, and other AI-facing metadata blocks. These envelopes must be
+ * removed BEFORE normalization, because `stripInboundMetadata` relies on
+ * newline structure and fenced `json` code fences to locate sentinels; once
+ * `normalizeSessionText` collapses newlines into spaces, stripping is
+ * impossible.
+ *
+ * See: https://github.com/openclaw/openclaw/issues/63921
+ */
+function stripInboundMetadataForUserRole(text: string, role: "user" | "assistant"): string {
+  if (role !== "user") {
+    return text;
+  }
+  return stripInboundMetadata(text);
+}
+
+const GENERATED_SYSTEM_MESSAGE_RE = /^System(?: \(untrusted\))?: \[[^\]]+\]\s*/;
+
+function isGeneratedSystemWrapperMessage(text: string, role: "user" | "assistant"): boolean {
+  if (role !== "user") {
+    return false;
+  }
+  return GENERATED_SYSTEM_MESSAGE_RE.test(text);
+}
+
+function isGeneratedCronPromptMessage(text: string, role: "user" | "assistant"): boolean {
+  if (role !== "user") {
+    return false;
+  }
+  return DIRECT_CRON_PROMPT_RE.test(text);
+}
+
+function isGeneratedHeartbeatPromptMessage(text: string, role: "user" | "assistant"): boolean {
+  return role === "user" && isHeartbeatUserMessage({ role, content: text }, HEARTBEAT_PROMPT);
+}
+
+function sanitizeSessionText(text: string, role: "user" | "assistant"): string | null {
+  const strippedInbound = stripInboundMetadataForUserRole(text, role);
+  const strippedInternal = stripInternalRuntimeContext(strippedInbound);
+  const normalized = normalizeSessionText(strippedInternal);
+  if (!normalized) {
+    return null;
+  }
+  if (isGeneratedSystemWrapperMessage(normalized, role)) {
+    return null;
+  }
+  if (isGeneratedCronPromptMessage(normalized, role)) {
+    return null;
+  }
+  if (isGeneratedHeartbeatPromptMessage(normalized, role)) {
+    return null;
+  }
+  if (isSilentReplyPayloadText(normalized)) {
+    return null;
+  }
+  // Assistant-side machinery acks: HEARTBEAT_OK is the canonical "all clear,
+  // nothing to do" reply to a heartbeat tick. Drop on the assistant side
+  // directly so we do not have to rely on cross-message coupling with the
+  // preceding user message (which a real user could spoof).
+  if (role === "assistant" && normalized === HEARTBEAT_TOKEN) {
+    return null;
+  }
+  const withoutSystemEnvelope = normalized.replace(GENERATED_SYSTEM_MESSAGE_RE, "").trim();
+  if (isExecCompletionEvent(withoutSystemEnvelope)) {
+    return null;
+  }
+  return normalized;
+}
+
+function parseSessionTimestampMs(
+  record: { timestamp?: unknown },
+  message: { timestamp?: unknown },
+): number {
+  const candidates = [message.timestamp, record.timestamp];
+  for (const value of candidates) {
+    if (typeof value === "number" && Number.isFinite(value)) {
+      const ms = value > 0 && value < 1e11 ? value * 1000 : value;
+      if (Number.isFinite(ms) && ms > 0 && ms <= MAX_DATE_TIMESTAMP_MS) {
+        return ms;
+      }
+    }
+    if (typeof value === "string") {
+      const parsed = Date.parse(value);
+      if (Number.isFinite(parsed) && parsed > 0) {
+        return parsed;
+      }
+    }
+  }
+  return 0;
+}
+
+function resolveSessionEntryParseYieldLines(opts: BuildSessionEntryOptions): number {
+  const configured = opts.parseYieldEveryLines;
+  if (typeof configured === "number" && Number.isFinite(configured)) {
+    return Math.max(1, Math.floor(configured));
+  }
+  return SESSION_ENTRY_PARSE_YIELD_LINES;
+}
+
+async function yieldSessionEntryParseIfNeeded(
+  lineIndex: number,
+  everyLines: number,
+): Promise<void> {
+  if (lineIndex > 0 && lineIndex % everyLines === 0) {
+    await new Promise<void>((resolve) => {
+      setImmediate(resolve);
+    });
+  }
+}
+
+export async function buildSessionEntry(
+  absPath: string,
+  opts: BuildSessionEntryOptions = {},
+): Promise<SessionFileEntry | null> {
+  try {
+    const regularFile = await statRegularFile(absPath);
+    if (regularFile.missing) {
+      return null;
+    }
+    const stat = regularFile.stat;
+    if (shouldSkipTranscriptFileForDreaming(absPath)) {
+      return {
+        path: sessionPathForFile(absPath),
+        absPath,
+        mtimeMs: stat.mtimeMs,
+        size: stat.size,
+        hash: hashText("\n\n"),
+        content: "",
+        lineMap: [],
+        messageTimestampsMs: [],
+      };
+    }
+    const raw = (
+      await retryTransientMemoryRead(
+        () => readRegularFile({ filePath: absPath }),
+        `read session transcript ${absPath}`,
+      )
+    ).buffer.toString("utf-8");
+    const collected: string[] = [];
+    const lineMap: number[] = [];
+    const messageTimestampsMs: number[] = [];
+    const parseYieldEveryLines = resolveSessionEntryParseYieldLines(opts);
+    const sessionStoreClassification =
+      opts.generatedByDreamingNarrative === undefined || opts.generatedByCronRun === undefined
+        ? classifySessionTranscriptFromSessionStore(absPath)
+        : null;
+    let generatedByDreamingNarrative =
+      opts.generatedByDreamingNarrative ??
+      sessionStoreClassification?.generatedByDreamingNarrative ??
+      false;
+    let generatedByCronRun =
+      opts.generatedByCronRun ?? sessionStoreClassification?.generatedByCronRun ?? false;
+    const allowArchiveContentCronClassification =
+      isUsageCountedSessionArchiveTranscriptPath(absPath);
+    for (let jsonlIdx = 0, lineStart = 0; lineStart <= raw.length; jsonlIdx++) {
+      await yieldSessionEntryParseIfNeeded(jsonlIdx, parseYieldEveryLines);
+      const newlineIndex = raw.indexOf("\n", lineStart);
+      const lineEnd = newlineIndex === -1 ? raw.length : newlineIndex;
+      const line = raw.slice(lineStart, lineEnd);
+      lineStart = newlineIndex === -1 ? raw.length + 1 : newlineIndex + 1;
+      if (!line.trim()) {
+        continue;
+      }
+      let record: unknown;
+      try {
+        record = JSON.parse(line);
+      } catch {
+        continue;
+      }
+      if (!generatedByDreamingNarrative && isDreamingNarrativeGeneratedRecord(record)) {
+        generatedByDreamingNarrative = true;
+      }
+      if (
+        !generatedByCronRun &&
+        allowArchiveContentCronClassification &&
+        isCronRunGeneratedRecord(record)
+      ) {
+        generatedByCronRun = true;
+        collected.length = 0;
+        lineMap.length = 0;
+        messageTimestampsMs.length = 0;
+      }
+      if (
+        !record ||
+        typeof record !== "object" ||
+        (record as { type?: unknown }).type !== "message"
+      ) {
+        continue;
+      }
+      const message = (record as { message?: unknown }).message as
+        | { role?: unknown; content?: unknown; provenance?: unknown }
+        | undefined;
+      if (!message || typeof message.role !== "string") {
+        continue;
+      }
+      if (message.role !== "user" && message.role !== "assistant") {
+        continue;
+      }
+      if (message.role === "user" && hasInterSessionUserProvenance(message)) {
+        continue;
+      }
+      const rawText = collectRawSessionText(message.content);
+      if (rawText === null) {
+        continue;
+      }
+      if (
+        !generatedByCronRun &&
+        allowArchiveContentCronClassification &&
+        isGeneratedCronPromptMessage(normalizeSessionText(rawText), message.role)
+      ) {
+        generatedByCronRun = true;
+        collected.length = 0;
+        lineMap.length = 0;
+        messageTimestampsMs.length = 0;
+      }
+      const text = sanitizeSessionText(rawText, message.role);
+      if (!text) {
+        // Assistant-side machinery (silent replies, system wrappers) is already
+        // dropped by sanitizeSessionText. We deliberately do NOT use the prior
+        // user message's pattern-match to drop the next assistant message:
+        // user-typed text can match those same patterns (`[cron:...]`,
+        // `System (untrusted): ...`) and a cross-message drop would let users
+        // exfiltrate real assistant replies from the dreaming corpus by
+        // prefixing their own prompt. See PR #70737 review (aisle-research-bot).
+        continue;
+      }
+      if (generatedByDreamingNarrative || generatedByCronRun) {
+        continue;
+      }
+      const safe = redactSensitiveText(text, { mode: "tools" });
+      const label = message.role === "user" ? "User" : "Assistant";
+      const renderedLines = renderSessionExportLines(label, safe);
+      const timestampMs = parseSessionTimestampMs(
+        record as { timestamp?: unknown },
+        message as { timestamp?: unknown },
+      );
+      collected.push(...renderedLines);
+      lineMap.push(...renderedLines.map(() => jsonlIdx + 1));
+      messageTimestampsMs.push(...renderedLines.map(() => timestampMs));
+    }
+    const content = collected.join("\n");
+    return {
+      path: sessionPathForFile(absPath),
+      absPath,
+      mtimeMs: stat.mtimeMs,
+      size: stat.size,
+      hash: hashText(content + "\n" + lineMap.join(",") + "\n" + messageTimestampsMs.join(",")),
+      content,
+      lineMap,
+      messageTimestampsMs,
+      ...(generatedByDreamingNarrative ? { generatedByDreamingNarrative: true } : {}),
+      ...(generatedByCronRun ? { generatedByCronRun: true } : {}),
+    };
+  } catch (err) {
+    void logSessionFileReadFailure(absPath, err);
+    return null;
+  }
+}

packages/memory-host-sdk/src/host/session-transcript-corpus.ts

@@ -0,0 +1,425 @@
+// Accessor-backed transcript corpus discovery for memory/QMD session indexing.
+import fsSync from "node:fs";
+import path from "node:path";
+import { normalizeAgentId } from "./config-utils.js";
+import {
+  isDreamingNarrativeSessionStoreKey,
+  extractAgentIdFromSessionsDir,
+  canonicalizeMainSessionAlias,
+  getRuntimeConfig,
+  isCronRunSessionKey,
+  isSessionArchiveArtifactName,
+  isUsageCountedSessionTranscriptFileName,
+  listSessionEntries,
+  parseUsageCountedSessionIdFromFileName,
+  resolveSessionAgentId,
+  resolveSessionFilePath,
+  resolveStorePath,
+  type SessionEntry,
+} from "./openclaw-runtime-session.js";
+
+export type SessionTranscriptCorpusArtifactKind =
+  | "active-session"
+  | "archive-artifact"
+  | "orphan-file-artifact";
+
+export type SessionTranscriptCorpusEntry = {
+  agentId: string;
+  sessionFile: string;
+  sessionId: string;
+  artifactKind: SessionTranscriptCorpusArtifactKind;
+  sessionKey?: string;
+  /** True when this transcript belongs to an internal dreaming narrative run. */
+  generatedByDreamingNarrative?: boolean;
+  /** True when this transcript belongs to an isolated cron run session. */
+  generatedByCronRun?: boolean;
+};
+
+type SessionEntrySummary = {
+  sessionKey: string;
+  entry: SessionEntry;
+};
+
+function isDreamingNarrativeSessionKeyLike(value: unknown): boolean {
+  return typeof value === "string" && isDreamingNarrativeSessionStoreKey(value);
+}
+
+function normalizeComparablePath(pathname: string): string {
+  const resolved = path.resolve(pathname);
+  return process.platform === "win32" ? resolved.toLowerCase() : resolved;
+}
+
+function normalizeRealComparablePath(pathname: string): string {
+  try {
+    return normalizeComparablePath(fsSync.realpathSync(pathname));
+  } catch {
+    try {
+      return normalizeComparablePath(
+        path.join(fsSync.realpathSync(path.dirname(pathname)), path.basename(pathname)),
+      );
+    } catch {
+      return normalizeComparablePath(pathname);
+    }
+  }
+}
+
+function rememberArtifactDir(dirs: Map<string, string>, dir: string): void {
+  dirs.set(normalizeRealComparablePath(dir), dir);
+}
+
+function extractAgentIdFromSessionPath(absPath: string): string | null {
+  const parts = path.normalize(path.resolve(absPath)).split(path.sep).filter(Boolean);
+  const sessionsIndex = parts.lastIndexOf("sessions");
+  if (sessionsIndex < 2 || parts[sessionsIndex - 2] !== "agents") {
+    return null;
+  }
+  return parts[sessionsIndex - 1] || null;
+}
+
+function resolveSessionStoreTranscriptCorpusPath(
+  agentId: string,
+  sessionsDir: string,
+  entry: { sessionFile?: unknown; sessionId?: unknown } | undefined,
+): string | null {
+  const sessionFile =
+    typeof entry?.sessionFile === "string" && entry.sessionFile.trim().length > 0
+      ? entry.sessionFile.trim()
+      : undefined;
+  const sessionId =
+    typeof entry?.sessionId === "string" && entry.sessionId.trim().length > 0
+      ? entry.sessionId.trim()
+      : sessionFile
+        ? parseUsageCountedSessionIdFromFileName(path.basename(sessionFile))
+        : null;
+  if (!sessionId) {
+    return null;
+  }
+  try {
+    if (!sessionFile) {
+      return resolveSessionFilePath(sessionId, undefined, { agentId, sessionsDir });
+    }
+    const resolved = resolveSessionFilePath(
+      sessionId,
+      { sessionFile },
+      {
+        agentId,
+        sessionsDir,
+      },
+    );
+    if (!path.isAbsolute(sessionFile)) {
+      const candidate = path.resolve(sessionsDir, sessionFile);
+      if (
+        normalizeComparablePath(path.dirname(candidate)) !== normalizeComparablePath(sessionsDir)
+      ) {
+        return null;
+      }
+      return normalizeRealComparablePath(resolved) === normalizeRealComparablePath(candidate)
+        ? candidate
+        : null;
+    }
+    const pathAgentId = extractAgentIdFromSessionPath(sessionFile);
+    if (pathAgentId && normalizeAgentId(pathAgentId) !== normalizeAgentId(agentId)) {
+      return null;
+    }
+    return normalizeRealComparablePath(resolved) === normalizeRealComparablePath(sessionFile)
+      ? sessionFile
+      : null;
+  } catch {
+    return null;
+  }
+}
+
+function classifySessionEntry(
+  sessionKey: string,
+  entry: SessionEntry,
+  cronGeneratedSessionKeys: ReadonlySet<string>,
+): {
+  generatedByDreamingNarrative: boolean;
+  generatedByCronRun: boolean;
+} {
+  return {
+    generatedByDreamingNarrative:
+      isDreamingNarrativeSessionStoreKey(sessionKey) ||
+      isDreamingNarrativeSessionKeyLike(entry.spawnedBy),
+    generatedByCronRun: cronGeneratedSessionKeys.has(sessionKey),
+  };
+}
+
+function readParentSessionKeys(entry: SessionEntry | undefined): string[] {
+  const keys = new Set<string>();
+  for (const value of [entry?.parentSessionKey, entry?.spawnedBy]) {
+    if (typeof value !== "string") {
+      continue;
+    }
+    const trimmed = value.trim();
+    if (trimmed) {
+      keys.add(trimmed);
+    }
+  }
+  return [...keys];
+}
+
+function collectCronGeneratedSessionKeys(
+  summaries: readonly SessionEntrySummary[],
+): ReadonlySet<string> {
+  // Build the cron-generated closure once so active entries and archive
+  // artifacts share the same lineage classification.
+  const entriesByKey = new Map(summaries.map((summary) => [summary.sessionKey, summary.entry]));
+  const cronGeneratedKeys = new Set<string>();
+  const cache = new Map<string, boolean>();
+  const resolving = new Set<string>();
+
+  const isCronGenerated = (sessionKey: string, entry: SessionEntry | undefined): boolean => {
+    if (isCronRunSessionKey(sessionKey)) {
+      cache.set(sessionKey, true);
+      cronGeneratedKeys.add(sessionKey);
+      return true;
+    }
+    const cached = cache.get(sessionKey);
+    if (cached !== undefined) {
+      return cached;
+    }
+    if (resolving.has(sessionKey)) {
+      return false;
+    }
+
+    resolving.add(sessionKey);
+    const generated = readParentSessionKeys(entry).some(
+      (parentKey) =>
+        // Parent rows can be pruned before child rows; a cron-shaped parent key
+        // still carries cron lineage without requiring a store entry.
+        isCronRunSessionKey(parentKey) || isCronGenerated(parentKey, entriesByKey.get(parentKey)),
+    );
+    resolving.delete(sessionKey);
+    cache.set(sessionKey, generated);
+    if (generated) {
+      cronGeneratedKeys.add(sessionKey);
+    }
+    return generated;
+  };
+
+  for (const summary of summaries) {
+    isCronGenerated(summary.sessionKey, summary.entry);
+  }
+  return cronGeneratedKeys;
+}
+
+function isRegularSessionTranscriptFile(absPath: string): boolean {
+  try {
+    return fsSync.lstatSync(absPath).isFile();
+  } catch {
+    return false;
+  }
+}
+
+function toSessionStoreCorpusEntry(
+  agentId: string,
+  sessionsDir: string,
+  summary: SessionEntrySummary,
+  cronGeneratedSessionKeys: ReadonlySet<string>,
+): SessionTranscriptCorpusEntry | null {
+  const sessionFile = resolveSessionStoreTranscriptCorpusPath(agentId, sessionsDir, summary.entry);
+  if (!sessionFile || !isUsageCountedSessionTranscriptFileName(path.basename(sessionFile))) {
+    return null;
+  }
+  const sessionId =
+    typeof summary.entry.sessionId === "string" && summary.entry.sessionId.trim()
+      ? summary.entry.sessionId.trim()
+      : parseUsageCountedSessionIdFromFileName(path.basename(sessionFile));
+  if (!sessionId) {
+    return null;
+  }
+  const sessionKey = summary.sessionKey.trim();
+  const classification = classifySessionEntry(
+    summary.sessionKey,
+    summary.entry,
+    cronGeneratedSessionKeys,
+  );
+  return {
+    agentId,
+    artifactKind: "active-session",
+    sessionFile,
+    sessionId,
+    ...(sessionKey ? { sessionKey } : {}),
+    ...(classification.generatedByDreamingNarrative ? { generatedByDreamingNarrative: true } : {}),
+    ...(classification.generatedByCronRun ? { generatedByCronRun: true } : {}),
+  };
+}
+
+function listSessionTranscriptArtifactFiles(sessionsDir: string): string[] {
+  try {
+    return fsSync
+      .readdirSync(sessionsDir, { withFileTypes: true })
+      .filter((entry) => entry.isFile())
+      .map((entry) => entry.name)
+      .filter((name) => isUsageCountedSessionTranscriptFileName(name))
+      .map((name) => path.join(sessionsDir, name));
+  } catch {
+    return [];
+  }
+}
+
+function classifyTranscriptArtifact(
+  artifactPath: string,
+  activeEntriesByPath: ReadonlyMap<string, SessionTranscriptCorpusEntry>,
+): {
+  generatedByDreamingNarrative: boolean;
+  generatedByCronRun: boolean;
+} {
+  const directEntry = activeEntriesByPath.get(normalizeRealComparablePath(artifactPath));
+  if (directEntry) {
+    return {
+      generatedByDreamingNarrative: directEntry.generatedByDreamingNarrative === true,
+      generatedByCronRun: directEntry.generatedByCronRun === true,
+    };
+  }
+  const sessionsDir = path.dirname(artifactPath);
+  const primarySessionId = parseUsageCountedSessionIdFromFileName(path.basename(artifactPath));
+  const primaryEntry =
+    primarySessionId && isSessionArchiveArtifactName(path.basename(artifactPath))
+      ? activeEntriesByPath.get(
+          normalizeRealComparablePath(path.join(sessionsDir, `${primarySessionId}.jsonl`)),
+        )
+      : undefined;
+  return {
+    generatedByDreamingNarrative: primaryEntry?.generatedByDreamingNarrative === true,
+    generatedByCronRun: primaryEntry?.generatedByCronRun === true,
+  };
+}
+
+function toArtifactCorpusEntry(
+  agentId: string,
+  artifactPath: string,
+  activeEntriesByPath: ReadonlyMap<string, SessionTranscriptCorpusEntry>,
+): SessionTranscriptCorpusEntry | null {
+  const sessionId = parseUsageCountedSessionIdFromFileName(path.basename(artifactPath));
+  if (!sessionId) {
+    return null;
+  }
+  const artifactKind = isSessionArchiveArtifactName(path.basename(artifactPath))
+    ? "archive-artifact"
+    : "orphan-file-artifact";
+  const classification = classifyTranscriptArtifact(artifactPath, activeEntriesByPath);
+  return {
+    agentId,
+    artifactKind,
+    sessionFile: artifactPath,
+    sessionId,
+    ...(classification.generatedByDreamingNarrative ? { generatedByDreamingNarrative: true } : {}),
+    ...(classification.generatedByCronRun ? { generatedByCronRun: true } : {}),
+  };
+}
+
+export function listSessionTranscriptCorpusEntriesForAgentSync(
+  agentId: string,
+): SessionTranscriptCorpusEntry[] {
+  const normalizedAgentId = normalizeAgentId(agentId);
+  const cfg = getRuntimeConfig();
+  const configuredStore = cfg.session?.store;
+  const storePath = resolveStorePath(configuredStore, {
+    agentId: normalizedAgentId,
+  });
+  const sessionsDir = path.dirname(storePath);
+  const fixedStoreOwnerAgentId = extractAgentIdFromSessionsDir(sessionsDir);
+  const isAgentOwnedFixedStore =
+    fixedStoreOwnerAgentId !== null &&
+    normalizeAgentId(fixedStoreOwnerAgentId) === normalizedAgentId;
+  const isSharedFixedStore =
+    typeof configuredStore === "string" &&
+    configuredStore.trim().length > 0 &&
+    !configuredStore.includes("{agentId}") &&
+    !isAgentOwnedFixedStore;
+  const activeEntriesByPath = new Map<string, SessionTranscriptCorpusEntry>();
+  const activeEntryOwnersByPath = new Map<string, string>();
+  const artifactDirsByPath = new Map<string, string>();
+  rememberArtifactDir(artifactDirsByPath, sessionsDir);
+  const sessionEntries = listSessionEntries({
+    agentId: normalizedAgentId,
+    hydrateSkillPromptRefs: false,
+    storePath,
+  });
+  const cronGeneratedSessionKeys = collectCronGeneratedSessionKeys(sessionEntries);
+  for (const summary of sessionEntries) {
+    const sessionKey = isSharedFixedStore
+      ? summary.sessionKey
+      : canonicalizeMainSessionAlias({
+          cfg,
+          agentId: normalizedAgentId,
+          sessionKey: summary.sessionKey,
+        });
+    const ownerAgentId = resolveSessionAgentId({
+      config: cfg,
+      sessionKey,
+      ...(isSharedFixedStore ? {} : { fallbackAgentId: normalizedAgentId }),
+    });
+    const entry = toSessionStoreCorpusEntry(
+      ownerAgentId,
+      sessionsDir,
+      summary,
+      cronGeneratedSessionKeys,
+    );
+    if (!entry) {
+      continue;
+    }
+    const normalizedEntryPath = normalizeRealComparablePath(entry.sessionFile);
+    activeEntryOwnersByPath.set(normalizedEntryPath, ownerAgentId);
+    rememberArtifactDir(artifactDirsByPath, path.dirname(entry.sessionFile));
+    if (ownerAgentId === normalizedAgentId) {
+      activeEntriesByPath.set(normalizedEntryPath, entry);
+    }
+  }
+  const includeUnownedArtifacts = !isSharedFixedStore;
+  const corpusEntries = [...activeEntriesByPath.values()].filter((entry) =>
+    isRegularSessionTranscriptFile(entry.sessionFile),
+  );
+  const scannedArtifactPaths = new Set<string>();
+  for (const artifactDir of artifactDirsByPath.values()) {
+    for (const artifactPath of listSessionTranscriptArtifactFiles(artifactDir)) {
+      const normalizedArtifactPath = normalizeRealComparablePath(artifactPath);
+      if (scannedArtifactPaths.has(normalizedArtifactPath)) {
+        continue;
+      }
+      scannedArtifactPaths.add(normalizedArtifactPath);
+      if (activeEntriesByPath.has(normalizedArtifactPath)) {
+        continue;
+      }
+      const artifactOwner = activeEntryOwnersByPath.get(normalizedArtifactPath);
+      if (artifactOwner) {
+        continue;
+      }
+      const primarySessionId = parseUsageCountedSessionIdFromFileName(path.basename(artifactPath));
+      const primaryOwner =
+        primarySessionId && isSessionArchiveArtifactName(path.basename(artifactPath))
+          ? activeEntryOwnersByPath.get(
+              normalizeRealComparablePath(
+                path.join(path.dirname(artifactPath), `${primarySessionId}.jsonl`),
+              ),
+            )
+          : undefined;
+      if (primaryOwner && primaryOwner !== normalizedAgentId) {
+        continue;
+      }
+      if (!primaryOwner && !includeUnownedArtifacts) {
+        continue;
+      }
+      const entry = toArtifactCorpusEntry(normalizedAgentId, artifactPath, activeEntriesByPath);
+      if (entry) {
+        corpusEntries.push(entry);
+      }
+    }
+  }
+  return corpusEntries;
+}
+
+/**
+ * Lists transcript corpus entries for QMD/memory indexing.
+ *
+ * Active sessions come from the session accessor seam; retained reset/delete
+ * transcript artifacts remain explicit file artifacts until core owns archive
+ * artifact enumeration.
+ */
+export async function listSessionTranscriptCorpusEntriesForAgent(
+  agentId: string,
+): Promise<SessionTranscriptCorpusEntry[]> {
+  return listSessionTranscriptCorpusEntriesForAgentSync(agentId);
+}

packages/memory-host-sdk/src/host/sqlite-vec-platform-variant.ts

@@ -0,0 +1,32 @@
+// Memory Host SDK module implements sqlite vec platform variant behavior.
+import { createRequire } from "node:module";
+
+// Resolves optional sqlite-vec native extension packages for the current platform.
+
+/** Package/file pair for one sqlite-vec native platform build. */
+type PlatformVariant = { readonly pkg: string; readonly file: string };
+
+const PLATFORM_VARIANTS: Readonly<Record<string, PlatformVariant | undefined>> = {
+  "linux-x64": { pkg: "sqlite-vec-linux-x64", file: "vec0.so" },
+  "linux-arm64": { pkg: "sqlite-vec-linux-arm64", file: "vec0.so" },
+  "darwin-x64": { pkg: "sqlite-vec-darwin-x64", file: "vec0.dylib" },
+  "darwin-arm64": { pkg: "sqlite-vec-darwin-arm64", file: "vec0.dylib" },
+  "win32-x64": { pkg: "sqlite-vec-windows-x64", file: "vec0.dll" },
+};
+
+/** Resolve the installed sqlite-vec native extension for the current platform if present. */
+export function resolveSqliteVecPlatformVariant():
+  | { pkg: string; extensionPath: string }
+  | undefined {
+  const entry = PLATFORM_VARIANTS[`${process.platform}-${process.arch}`];
+  if (!entry) {
+    return undefined;
+  }
+  try {
+    const requireForResolve = createRequire(import.meta.url);
+    const extensionPath = requireForResolve.resolve(`${entry.pkg}/${entry.file}`);
+    return { pkg: entry.pkg, extensionPath };
+  } catch {
+    return undefined;
+  }
+}

packages/memory-host-sdk/src/host/sqlite-vec.test.ts

@@ -0,0 +1,258 @@
+// Memory Host SDK tests cover sqlite vec behavior.
+import { existsSync } from "node:fs";
+import { createRequire } from "node:module";
+import { afterEach, describe, expect, it, vi } from "vitest";
+
+function mockMissingSqliteVecPackage(): void {
+  vi.doMock("sqlite-vec", () => {
+    const err = new Error("Cannot find package 'sqlite-vec' imported from sqlite-vec.test.ts");
+    Object.assign(err, { code: "ERR_MODULE_NOT_FOUND" });
+    throw err;
+  });
+}
+
+function mockFailingSqliteVecPackage(): void {
+  vi.doMock("sqlite-vec", () => ({
+    getLoadablePath: () => "/install/node_modules/sqlite-vec-linux-x64/vec0.so",
+    load: () => {
+      throw new Error("bundled sqlite-vec load failed");
+    },
+  }));
+}
+
+function mockPlatformVariantResolver(
+  value: { pkg: string; extensionPath: string } | undefined,
+): void {
+  vi.doMock("./sqlite-vec-platform-variant.js", () => ({
+    resolveSqliteVecPlatformVariant: () => value,
+  }));
+}
+
+async function importLoader() {
+  return import("./sqlite-vec.js");
+}
+
+function createDbMock(params?: { readonly healthError?: Error }) {
+  const get = vi.fn(() => ({ version: "v0.1.9" }));
+  const prepare = vi.fn(() => {
+    if (params?.healthError) {
+      throw params.healthError;
+    }
+    return { get };
+  });
+  return {
+    db: {
+      enableLoadExtension: vi.fn(),
+      loadExtension: vi.fn(),
+      prepare,
+    },
+    get,
+    prepare,
+  };
+}
+
+afterEach(() => {
+  vi.doUnmock("sqlite-vec");
+  vi.doUnmock("./sqlite-vec-platform-variant.js");
+  vi.resetModules();
+});
+
+const CURRENT_PLATFORM_VARIANTS: Readonly<
+  Record<string, { readonly pkg: string; readonly file: string } | undefined>
+> = {
+  "linux-x64": { pkg: "sqlite-vec-linux-x64", file: "vec0.so" },
+  "linux-arm64": { pkg: "sqlite-vec-linux-arm64", file: "vec0.so" },
+  "darwin-x64": { pkg: "sqlite-vec-darwin-x64", file: "vec0.dylib" },
+  "darwin-arm64": { pkg: "sqlite-vec-darwin-arm64", file: "vec0.dylib" },
+  "win32-x64": { pkg: "sqlite-vec-windows-x64", file: "vec0.dll" },
+};
+
+function isMissingModuleError(err: unknown): boolean {
+  const code =
+    err && typeof err === "object" && "code" in err ? (err as { code?: unknown }).code : undefined;
+  return code === "MODULE_NOT_FOUND" || code === "ERR_MODULE_NOT_FOUND";
+}
+
+describe("loadSqliteVecExtension", () => {
+  it("loads explicit extensionPath without importing bundled sqlite-vec", async () => {
+    mockMissingSqliteVecPackage();
+    const { loadSqliteVecExtension } = await importLoader();
+    const { db, prepare } = createDbMock();
+
+    await expect(
+      loadSqliteVecExtension({
+        db: db as never,
+        extensionPath: "/opt/openclaw/sqlite-vec.so",
+      }),
+    ).resolves.toEqual({ ok: true, extensionPath: "/opt/openclaw/sqlite-vec.so" });
+    expect(db.enableLoadExtension).toHaveBeenCalledWith(true);
+    expect(db.loadExtension).toHaveBeenCalledWith("/opt/openclaw/sqlite-vec.so");
+    expect(prepare).toHaveBeenCalledWith("SELECT vec_version() AS version");
+  });
+
+  it("rejects a loaded extension when sqlite-vec functions are unavailable", async () => {
+    mockMissingSqliteVecPackage();
+    const { loadSqliteVecExtension } = await importLoader();
+    const { db } = createDbMock({ healthError: new Error("no such function: vec_version") });
+
+    const result = await loadSqliteVecExtension({
+      db: db as never,
+      extensionPath: "/opt/openclaw/sqlite-vec.so",
+    });
+
+    expect(result).toEqual({
+      ok: false,
+      error:
+        "sqlite-vec health check failed after loading /opt/openclaw/sqlite-vec.so | no such function: vec_version",
+    });
+    expect(db.loadExtension).toHaveBeenCalledWith("/opt/openclaw/sqlite-vec.so");
+  });
+
+  it("returns a valid memorySearch extensionPath hint when sqlite-vec is absent", async () => {
+    mockMissingSqliteVecPackage();
+    mockPlatformVariantResolver(undefined);
+    const { loadSqliteVecExtension } = await importLoader();
+    const { db } = createDbMock();
+
+    const result = await loadSqliteVecExtension({ db: db as never });
+
+    expect(result).toEqual({
+      ok: false,
+      error: expect.stringMatching(
+        /^sqlite-vec package is not installed\. Set agents\.defaults\.memorySearch\.store\.vector\.extensionPath, or an agent-specific memorySearch\.store\.vector\.extensionPath, to a sqlite-vec loadable extension path\. Original error: (?:\[vitest\] There was an error when mocking a module\. If you are using "vi\.mock" factory, make sure there are no top level variables inside, since this call is hoisted to top of the file\. Read more: https:\/\/vitest\.dev\/api\/vi\.html#vi-mock \| )?Cannot find package 'sqlite-vec' imported from sqlite-vec\.test\.ts$/u,
+      ),
+    });
+    expect(result.error).not.toContain("memory.store.vector.extensionPath");
+    expect(db.enableLoadExtension).toHaveBeenCalledWith(true);
+    expect(db.loadExtension).not.toHaveBeenCalled();
+  });
+
+  it("falls back to the platform-specific sqlite-vec variant when only that package is installed", async () => {
+    mockMissingSqliteVecPackage();
+    mockPlatformVariantResolver({
+      pkg: "sqlite-vec-linux-x64",
+      extensionPath: "/install/node_modules/sqlite-vec-linux-x64/vec0.so",
+    });
+    const { loadSqliteVecExtension } = await importLoader();
+    const { db, prepare } = createDbMock();
+
+    const result = await loadSqliteVecExtension({ db: db as never });
+
+    expect(result).toEqual({
+      ok: true,
+      extensionPath: "/install/node_modules/sqlite-vec-linux-x64/vec0.so",
+    });
+    expect(db.enableLoadExtension).toHaveBeenCalledWith(true);
+    expect(db.loadExtension).toHaveBeenCalledWith(
+      "/install/node_modules/sqlite-vec-linux-x64/vec0.so",
+    );
+    expect(prepare).toHaveBeenCalledWith("SELECT vec_version() AS version");
+  });
+
+  it("falls back to the platform variant when bundled sqlite-vec load fails", async () => {
+    mockFailingSqliteVecPackage();
+    mockPlatformVariantResolver({
+      pkg: "sqlite-vec-linux-x64",
+      extensionPath: "/install/node_modules/sqlite-vec-linux-x64/vec0.so",
+    });
+    const { loadSqliteVecExtension } = await importLoader();
+    const { db, prepare } = createDbMock();
+
+    const result = await loadSqliteVecExtension({ db: db as never });
+
+    expect(result).toEqual({
+      ok: true,
+      extensionPath: "/install/node_modules/sqlite-vec-linux-x64/vec0.so",
+    });
+    expect(db.loadExtension).toHaveBeenCalledWith(
+      "/install/node_modules/sqlite-vec-linux-x64/vec0.so",
+    );
+    expect(prepare).toHaveBeenCalledWith("SELECT vec_version() AS version");
+  });
+
+  it("resolves the installed platform variant through its exported vec0 subpath", async () => {
+    const entry = CURRENT_PLATFORM_VARIANTS[`${process.platform}-${process.arch}`];
+    if (!entry) {
+      return;
+    }
+
+    const requireForResolve = createRequire(import.meta.url);
+    let expectedPath: string;
+    try {
+      expectedPath = requireForResolve.resolve(`${entry.pkg}/${entry.file}`);
+    } catch (err) {
+      if (isMissingModuleError(err)) {
+        return;
+      }
+      throw err;
+    }
+
+    const { resolveSqliteVecPlatformVariant } = await import("./sqlite-vec-platform-variant.js");
+
+    expect(resolveSqliteVecPlatformVariant()).toEqual({
+      pkg: entry.pkg,
+      extensionPath: expectedPath,
+    });
+    expect(existsSync(expectedPath)).toBe(true);
+  });
+
+  it("preserves the extensionPath config hint when the platform variant loadExtension call throws", async () => {
+    mockMissingSqliteVecPackage();
+    mockPlatformVariantResolver({
+      pkg: "sqlite-vec-linux-x64",
+      extensionPath: "/install/node_modules/sqlite-vec-linux-x64/vec0.so",
+    });
+    const { loadSqliteVecExtension } = await importLoader();
+    const { db } = createDbMock();
+    db.loadExtension.mockImplementation(() => {
+      throw new Error("dlopen failed: file not found");
+    });
+
+    const result = await loadSqliteVecExtension({ db: db as never });
+
+    expect(result).toEqual({
+      ok: false,
+      error:
+        "sqlite-vec platform variant sqlite-vec-linux-x64 failed to load from /install/node_modules/sqlite-vec-linux-x64/vec0.so. Set agents.defaults.memorySearch.store.vector.extensionPath, or an agent-specific memorySearch.store.vector.extensionPath, to a sqlite-vec loadable extension path. Original error: dlopen failed: file not found",
+    });
+  });
+
+  it("rejects a platform variant when sqlite-vec functions are unavailable", async () => {
+    mockMissingSqliteVecPackage();
+    mockPlatformVariantResolver({
+      pkg: "sqlite-vec-linux-x64",
+      extensionPath: "/install/node_modules/sqlite-vec-linux-x64/vec0.so",
+    });
+    const { loadSqliteVecExtension } = await importLoader();
+    const { db } = createDbMock({ healthError: new Error("no such function: vec_version") });
+
+    const result = await loadSqliteVecExtension({ db: db as never });
+
+    expect(result).toEqual({
+      ok: false,
+      error:
+        "sqlite-vec platform variant sqlite-vec-linux-x64 failed to load from /install/node_modules/sqlite-vec-linux-x64/vec0.so. Set agents.defaults.memorySearch.store.vector.extensionPath, or an agent-specific memorySearch.store.vector.extensionPath, to a sqlite-vec loadable extension path. Original error: sqlite-vec health check failed after loading /install/node_modules/sqlite-vec-linux-x64/vec0.so | no such function: vec_version",
+    });
+  });
+
+  it("preserves bundled sqlite-vec and platform variant errors when both fail", async () => {
+    mockFailingSqliteVecPackage();
+    mockPlatformVariantResolver({
+      pkg: "sqlite-vec-linux-x64",
+      extensionPath: "/install/node_modules/sqlite-vec-linux-x64/vec0.so",
+    });
+    const { loadSqliteVecExtension } = await importLoader();
+    const { db } = createDbMock();
+    db.loadExtension.mockImplementation(() => {
+      throw new Error("platform variant failed");
+    });
+
+    const result = await loadSqliteVecExtension({ db: db as never });
+
+    expect(result).toEqual({
+      ok: false,
+      error:
+        "sqlite-vec package failed to load, and platform variant sqlite-vec-linux-x64 failed to load from /install/node_modules/sqlite-vec-linux-x64/vec0.so. Set agents.defaults.memorySearch.store.vector.extensionPath, or an agent-specific memorySearch.store.vector.extensionPath, to a sqlite-vec loadable extension path. Package error: bundled sqlite-vec load failed. Variant error: platform variant failed",
+    });
+  });
+});

packages/memory-host-sdk/src/host/sqlite-vec.ts

@@ -0,0 +1,105 @@
+// Memory Host SDK module implements sqlite vec behavior.
+import type { DatabaseSync } from "node:sqlite";
+import { formatErrorMessage } from "./error-utils.js";
+import { resolveSqliteVecPlatformVariant } from "./sqlite-vec-platform-variant.js";
+import { normalizeOptionalString } from "./string-utils.js";
+
+type SqliteVecModule = {
+  getLoadablePath: () => string;
+  load: (db: DatabaseSync) => void;
+};
+
+const SQLITE_VEC_MODULE_ID = "sqlite-vec";
+const SQLITE_VEC_CONFIG_HINT =
+  "Set agents.defaults.memorySearch.store.vector.extensionPath, or an agent-specific memorySearch.store.vector.extensionPath, to a sqlite-vec loadable extension path.";
+
+async function loadSqliteVecModule(): Promise<SqliteVecModule> {
+  return import(SQLITE_VEC_MODULE_ID) as Promise<SqliteVecModule>;
+}
+
+function isMissingSqliteVecPackageError(err: unknown): boolean {
+  const message = formatErrorMessage(err);
+  const code =
+    err && typeof err === "object" && "code" in err ? (err as { code?: unknown }).code : undefined;
+  const missingSqliteVec = /Cannot find (?:package|module) ['"]sqlite-vec['"]/u.test(message);
+  return (
+    missingSqliteVec &&
+    (code === undefined || code === "ERR_MODULE_NOT_FOUND" || code === "MODULE_NOT_FOUND")
+  );
+}
+
+function assertSqliteVecAvailable(db: DatabaseSync, source: string): void {
+  try {
+    const row = db.prepare("SELECT vec_version() AS version").get() as
+      | { version?: unknown }
+      | undefined;
+    if (typeof row?.version !== "string" || row.version.trim().length === 0) {
+      throw new Error("vec_version() did not return a version");
+    }
+  } catch (err) {
+    throw new Error(`sqlite-vec health check failed after loading ${source}`, { cause: err });
+  }
+}
+
+function loadExtensionAndVerify(db: DatabaseSync, extensionPath: string): void {
+  db.loadExtension(extensionPath);
+  assertSqliteVecAvailable(db, extensionPath);
+}
+
+export async function loadSqliteVecExtension(params: {
+  db: DatabaseSync;
+  extensionPath?: string;
+}): Promise<{ ok: boolean; extensionPath?: string; error?: string }> {
+  try {
+    const resolvedPath = normalizeOptionalString(params.extensionPath);
+    params.db.enableLoadExtension(true);
+    if (resolvedPath) {
+      loadExtensionAndVerify(params.db, resolvedPath);
+      return { ok: true, extensionPath: resolvedPath };
+    }
+
+    try {
+      const sqliteVec = await loadSqliteVecModule();
+      const extensionPath = sqliteVec.getLoadablePath();
+      sqliteVec.load(params.db);
+      assertSqliteVecAvailable(params.db, extensionPath);
+      return { ok: true, extensionPath };
+    } catch (err) {
+      // Optional-dep installs sometimes land only the platform-specific variant
+      // (e.g. sqlite-vec-linux-x64) without the meta sqlite-vec package. Load
+      // the loadable extension straight from the variant when we can find it.
+      // Bundled runtimes can also fail the meta-package import while the native
+      // variant is still present, so try the concrete extension before failing.
+      const variant = resolveSqliteVecPlatformVariant();
+      if (!variant) {
+        if (!isMissingSqliteVecPackageError(err)) {
+          throw err;
+        }
+        const message = formatErrorMessage(err);
+        return {
+          ok: false,
+          error: `sqlite-vec package is not installed. ${SQLITE_VEC_CONFIG_HINT} Original error: ${message}`,
+        };
+      }
+      try {
+        loadExtensionAndVerify(params.db, variant.extensionPath);
+        return { ok: true, extensionPath: variant.extensionPath };
+      } catch (variantErr) {
+        const message = formatErrorMessage(variantErr);
+        if (!isMissingSqliteVecPackageError(err)) {
+          const packageMessage = formatErrorMessage(err);
+          return {
+            ok: false,
+            error: `sqlite-vec package failed to load, and platform variant ${variant.pkg} failed to load from ${variant.extensionPath}. ${SQLITE_VEC_CONFIG_HINT} Package error: ${packageMessage}. Variant error: ${message}`,
+          };
+        }
+        return {
+          ok: false,
+          error: `sqlite-vec platform variant ${variant.pkg} failed to load from ${variant.extensionPath}. ${SQLITE_VEC_CONFIG_HINT} Original error: ${message}`,
+        };
+      }
+    }
+  } catch (err) {
+    return { ok: false, error: formatErrorMessage(err) };
+  }
+}

packages/memory-host-sdk/src/host/sqlite-wal.ts

@@ -0,0 +1,14 @@
+// Public SQLite WAL maintenance facade for memory database callers.
+
+export {
+  DEFAULT_SQLITE_WAL_AUTOCHECKPOINT_PAGES,
+  DEFAULT_SQLITE_WAL_CHECKPOINT_INTERVAL_MS,
+  DEFAULT_SQLITE_WAL_TRUNCATE_INTERVAL_MS,
+  configureSqliteConnectionPragmas,
+  configureSqliteWalMaintenance,
+} from "./openclaw-runtime-io.js";
+export type {
+  SqliteConnectionPragmaOptions,
+  SqliteWalMaintenance,
+  SqliteWalMaintenanceOptions,
+} from "./openclaw-runtime-io.js";

packages/memory-host-sdk/src/host/sqlite.ts

@@ -0,0 +1,55 @@
+// Memory Host SDK module implements sqlite behavior.
+import { createRequire } from "node:module";
+import type { DatabaseSync } from "node:sqlite";
+import { formatErrorMessage } from "./error-utils.js";
+import {
+  configureSqliteConnectionPragmas,
+  configureSqliteWalMaintenance,
+  type SqliteConnectionPragmaOptions,
+  type SqliteWalMaintenance,
+  type SqliteWalMaintenanceOptions,
+} from "./sqlite-wal.js";
+import { installProcessWarningFilter } from "./warning-filter.js";
+
+const require = createRequire(import.meta.url);
+const sqliteWalMaintenanceByDb = new WeakMap<DatabaseSync, SqliteWalMaintenance>();
+
+export function requireNodeSqlite(): typeof import("node:sqlite") {
+  installProcessWarningFilter();
+  try {
+    return require("node:sqlite") as typeof import("node:sqlite");
+  } catch (err) {
+    const message = formatErrorMessage(err);
+    // Node distributions can ship without the experimental builtin SQLite module.
+    // Surface an actionable error instead of the generic "unknown builtin module".
+    throw new Error(
+      `SQLite support is unavailable in this Node runtime (missing node:sqlite). ${message}`,
+      { cause: err },
+    );
+  }
+}
+
+export function configureMemorySqliteWalMaintenance(
+  db: DatabaseSync,
+  options?: SqliteWalMaintenanceOptions & Pick<SqliteConnectionPragmaOptions, "busyTimeoutMs">,
+): SqliteWalMaintenance {
+  const existing = sqliteWalMaintenanceByDb.get(db);
+  if (existing) {
+    return existing;
+  }
+  const maintenance =
+    options?.busyTimeoutMs === undefined
+      ? configureSqliteWalMaintenance(db, options)
+      : configureSqliteConnectionPragmas(db, options);
+  sqliteWalMaintenanceByDb.set(db, maintenance);
+  return maintenance;
+}
+
+export function closeMemorySqliteWalMaintenance(db: DatabaseSync): boolean {
+  const maintenance = sqliteWalMaintenanceByDb.get(db);
+  if (!maintenance) {
+    return true;
+  }
+  sqliteWalMaintenanceByDb.delete(db);
+  return maintenance.close();
+}

packages/memory-host-sdk/src/host/ssrf-policy.ts

@@ -0,0 +1,11 @@
+// Public SSRF policy shape accepted by memory host remote HTTP helpers.
+
+/** Host/network allowlist policy forwarded to the runtime SSRF guard. */
+export type SsrFPolicy = {
+  allowPrivateNetwork?: boolean;
+  dangerouslyAllowPrivateNetwork?: boolean;
+  allowRfc2544BenchmarkRange?: boolean;
+  allowIpv6UniqueLocalRange?: boolean;
+  allowedHostnames?: string[];
+  hostnameAllowlist?: string[];
+};

packages/memory-host-sdk/src/host/status-format.ts

@@ -0,0 +1,44 @@
+// Shared status text/tone formatter for memory health summaries.
+
+/** Display tone used by memory status renderers. */
+export type Tone = "ok" | "warn" | "muted";
+
+/** Resolve vector indexing state from enabled and availability flags. */
+export function resolveMemoryVectorState(vector: { enabled: boolean; available?: boolean }): {
+  tone: Tone;
+  state: "ready" | "unavailable" | "disabled" | "unknown";
+} {
+  if (!vector.enabled) {
+    return { tone: "muted", state: "disabled" };
+  }
+  if (vector.available === true) {
+    return { tone: "ok", state: "ready" };
+  }
+  if (vector.available === false) {
+    return { tone: "warn", state: "unavailable" };
+  }
+  return { tone: "muted", state: "unknown" };
+}
+
+/** Resolve full-text search state from enabled and availability flags. */
+export function resolveMemoryFtsState(fts: { enabled: boolean; available: boolean }): {
+  tone: Tone;
+  state: "ready" | "unavailable" | "disabled";
+} {
+  if (!fts.enabled) {
+    return { tone: "muted", state: "disabled" };
+  }
+  return fts.available ? { tone: "ok", state: "ready" } : { tone: "warn", state: "unavailable" };
+}
+
+/** Format cache state as concise status text with optional entry count. */
+export function resolveMemoryCacheSummary(cache: { enabled: boolean; entries?: number }): {
+  tone: Tone;
+  text: string;
+} {
+  if (!cache.enabled) {
+    return { tone: "muted", text: "cache off" };
+  }
+  const suffix = typeof cache.entries === "number" ? ` (${cache.entries})` : "";
+  return { tone: "ok", text: `cache on${suffix}` };
+}

packages/memory-host-sdk/src/host/string-utils.ts

@@ -0,0 +1,35 @@
+// Small string normalization helpers kept local to memory-host-sdk for package
+// builds that should not depend on the full normalization package graph.
+/** Normalize a non-empty string or return null. */
+export function normalizeNullableString(value: unknown): string | null {
+  if (typeof value !== "string") {
+    return null;
+  }
+  const trimmed = value.trim();
+  return trimmed ? trimmed : null;
+}
+
+/** Normalize a non-empty string or return undefined. */
+export function normalizeOptionalString(value: unknown): string | undefined {
+  return normalizeNullableString(value) ?? undefined;
+}
+
+/** Normalize a non-empty string to lowercase or return undefined. */
+export function normalizeOptionalLowercaseString(value: unknown): string | undefined {
+  return normalizeOptionalString(value)?.toLowerCase();
+}
+
+/** Normalize a value to lowercase text, defaulting to an empty string. */
+export function normalizeLowercaseStringOrEmpty(value: unknown): string {
+  return normalizeOptionalLowercaseString(value) ?? "";
+}
+
+/** Normalize an array-like list of values into non-empty strings. */
+export function normalizeStringEntries(values: ReadonlyArray<unknown>): string[] {
+  return values.map((value) => normalizeOptionalString(String(value)) ?? "").filter(Boolean);
+}
+
+/** Return unique strings preserving first-seen order. */
+export function uniqueStrings(values: Iterable<string>): string[] {
+  return [...new Set(values)];
+}