@ppagent/memory 0.1.4 → 0.4.0

package/llms.txt

@@ -1,6 +1,6 @@
 # @ppagent/memory
 
-> `@ppagent/memory` is an independent TypeScript/Node.js long-term memory package for AI agents. It stores conversation memory in LanceDB, extracts and searches entity relationships with Grafeo, supports user/chat/session scoped facts, compresses chat history into multi-resolution topics, and ingests Markdown documents as a searchable knowledge base.
+> `@ppagent/memory` is an independent TypeScript/Node.js long-term memory package for AI agents. It stores complete raw conversation messages in a pluggable vector store (LanceDB or SQLite/sqlite-vec, auto-detected per platform), builds importance-aware Topic summaries, returns fixed-budget Topics plus recent uncompressed messages as one context window, extracts and searches entity relationships with Grafeo, supports cached user/chat facts, and ingests Markdown documents as a searchable knowledge base.
 
 This file is written for AI coding agents and assistants. Use it as the primary package guide when generating code that depends on `@ppagent/memory`.
 
@@ -13,8 +13,9 @@ This file is written for AI coding agents and assistants. Use it as the primary
 - Source entry during local workspace development: `src/index.ts`
 - Runtime target: Node.js with global `fetch`
 - Storage engines:
-  - LanceDB for messages, topics, facts, sessions, documents, and chunks
+  - Vector store (dual backend, auto-detected): LanceDB or SQLite (better-sqlite3 + sqlite-vec + FTS5 with jieba Chinese tokenization) for messages, topics, facts, sessions, documents, and chunks. LanceDB requires native prebuilt binaries (unavailable on Intel macOS since 0.23.0); SQLite works everywhere and is the automatic fallback. RRF hybrid-search fusion runs in the shared domain layer, so search behavior is identical across backends.
   - Grafeo for entity/relation graph storage and graph search
+- Runtime configuration principle: `MemoryManager`/`MemoryStore` never read environment variables, config files, or implicit global state; every parameter enters through `MemoryConfig`. The separate manual migration CLI reads only the file explicitly supplied with `--config`.
 - LLM APIs: OpenAI-compatible `/chat/completions`
 - Embedding APIs: OpenAI-compatible `/embeddings`
 
@@ -48,7 +49,7 @@ npm install @ppagent/memory
 ```ts
 import {
   MemoryManager,
-  LanceService,
+  MemoryStore,
   DEFAULT_NODE_TYPES,
   DEFAULT_RELATION_TYPES,
   KIND_CONVERSATION,
@@ -62,11 +63,12 @@ import type {
   SearchResult,
   AddDocumentOptions,
   KnowledgeSearchOptions,
-  KnowledgeSearchResult,
-} from "@ppagent/memory";
+  KnowledgeSearchResult,
+  MemoryContextWindow,
+} from "@ppagent/memory";
 ```
 
-Most consumers should only instantiate `MemoryManager`. `LanceService` is exported for advanced storage tests or custom infrastructure work.
+Most consumers should only instantiate `MemoryManager`. `MemoryStore` (plus the `VectorStoreProvider` / `Filter` types) is exported for advanced storage tests or custom infrastructure work. Breaking change in 0.2.0: `LanceService` was replaced by `MemoryStore` + provider abstraction; string SQL filters were replaced by structured `Filter` objects.
 
 ## Minimal Example
 
@@ -120,7 +122,7 @@ const answer = await memory.ask({
 
 console.log(answer);
 
-memory.destroy();
+await memory.destroy();
 ```
 
 ## Complete Example
@@ -147,16 +149,19 @@ const memory = new MemoryManager({
   embeddingBatchSize: 20,
   embeddingConcurrency: 2,
 
-  sessionTokenLimit: 16_386,
-  historyWindowTokenLimit: 10_240,
-  topicRatio: [1, 5, 20],
-  detailMaxTokens: 2048,
-  summaryMaxTokens: 512,
-  conciseMaxTokens: 128,
-  maxConcurrentCompressions: 3,
+  compressedContextTokenLimit: 16 * 1024,
+  contextUsageRatio: 0.75,
+  precompressionRatio: 0.75,
+  compressionBatchRatio: 0.5,
+  compressionBatchTokenLimit: 0,
+  topicSummaryMaxTokens: 2048,
+  defaultModelContextTokens: 256 * 1024,
+  maxHistoryAgeMs: 0,
+  sessionIdleTtlMs: 30 * 60_000,
+  sessionSweepIntervalMs: 60_000,
+  maxConcurrentCompressions: 3,
   entitySimilarityThreshold: 0.92,
   defaultSearchLimit: 10,
-  recallBoostMs: 3_600_000,
 
   chunkStrategy: "markdown-heading",
   chunkMaxTokens: 800,
@@ -240,7 +245,7 @@ ChatMe 使用 TypeScript 构建，后端包含长期记忆系统。
 
 ## Memory
 
-长期记忆系统使用 LanceDB 存储向量，使用 Grafeo 存储知识图谱。`,
+长期记忆系统使用向量存储（LanceDB 或 SQLite，自动探测）存储向量，使用 Grafeo 存储知识图谱。`,
   userId,
   chatId,
   sessionId,
@@ -270,7 +275,7 @@ const answer = await memory.ask({
 
 console.log(answer);
 
-memory.destroy();
+await memory.destroy();
 ```
 
 ## Configuration Reference
@@ -279,9 +284,25 @@ memory.destroy();
 
 Required storage fields:
 
-- `lancedbPath: string` - local LanceDB directory.
+- `lancedbPath: string` - local LanceDB directory (used when the lancedb backend is active; its parent directory also hosts the backend marker file `.memory-provider`).
 - `grafeoPath: string` - Grafeo database path.
 
+Optional storage fields:
+
+- `provider?: "auto" | "lancedb" | "sqlite"` - vector store backend. Default `"auto"`: probes the @lancedb/lancedb native binding at init and falls back to sqlite when unavailable (e.g. Intel macOS, musl without prebuilds). The first resolved backend is persisted in a marker file; auto mode never silently switches an existing data directory to a different backend (it throws with migration guidance instead).
+- `sqlitePath?: string` - SQLite database file for the sqlite backend. Default: `path.join(path.dirname(lancedbPath), "memory.sqlite3")`, i.e. a `memory.sqlite3` file in the parent directory of `lancedbPath`. Parent directories are created automatically on init.
+
+Storage layout example with `lancedbPath: "./data/memory/lance"` and defaults:
+
+```
+data/memory/
+├── .memory-provider   # backend marker written on first init ("lancedb" or "sqlite")
+├── lance/             # LanceDB tables (lancedb backend only)
+└── memory.sqlite3     # SQLite database (sqlite backend only; plus -wal/-shm files while open)
+```
+
+Backend data files are not interchangeable; switching backends requires an explicit export/import migration.
+
 Required LLM fields:
 
 - `llmBaseUrl: string` - OpenAI-compatible base URL, for example `https://api.openai.com/v1`.
@@ -293,7 +314,7 @@ Required embedding fields:
 - `embeddingBaseUrl: string` - OpenAI-compatible embedding base URL.
 - `embeddingApiKey: string` - embedding API key.
 - `embeddingModel: string` - embedding model name.
-- `embeddingDimension: number` - vector dimension. This must match the embedding model and existing LanceDB schema.
+- `embeddingDimension: number` - vector dimension. This must match the embedding model and the existing vector store schema.
 
 Embedding fallback behavior:
 
@@ -308,18 +329,23 @@ HTTP and retry fields:
 - `embeddingBatchSize?: number` - max texts per embedding request. Default: `20`.
 - `embeddingConcurrency?: number` - concurrent embedding batches. Default: `2`.
 
-Conversation memory fields:
-
-- `sessionTokenLimit?: number` - when a session cache reaches this token count, compression is triggered. Default: `16386`.
-- `historyWindowTokenLimit?: number` - maximum token budget for the rebuilt compressed history window. Default: `10240`.
-- `topicRatio?: [number, number, number]` - number ratio for detail/summary/concise topics in the history window. Default: `[1, 5, 20]`.
-- `detailMaxTokens?: number` - max detailed summary tokens requested from the LLM. Default: `2048`.
-- `summaryMaxTokens?: number` - max medium summary tokens requested from the LLM. Default: `512`.
-- `conciseMaxTokens?: number` - max concise summary tokens requested from the LLM. Default: `128`.
-- `maxConcurrentCompressions?: number` - global semaphore limit for concurrent compression and knowledge ingestion tasks. Default: `3`.
-- `entitySimilarityThreshold?: number` - graph entity similarity threshold. Default: `0.92`.
-- `defaultSearchLimit?: number` - default `search` result limit. Default: `10`.
-- `recallBoostMs?: number` - each topic recall count acts like this many milliseconds of freshness boost. Default: `3_600_000`.
+Conversation memory fields:
+
+- `compressedContextTokenLimit?: number` - fixed token budget for compressed Topics. Default: `16384`; startup logs a warning above `32768`.
+- `contextUsageRatio?: number` - fraction of the current model context available to conversation history. Default: `0.75`.
+- `precompressionRatio?: number` - start background raw-message compression when raw usage reaches this fraction of its dynamic budget. Default: `0.75`.
+- `compressionBatchRatio?: number` - target fraction of the raw-message budget compressed in one batch. Default: `0.5`.
+- `compressionBatchTokenLimit?: number` - optional hard limit for one compression batch. Default: `0` (ratio only).
+- `topicSummaryMaxTokens?: number` - hard maximum for one Topic summary. The LLM uses less for low-value chat and more for facts, experience, decisions, preferences, constraints, and reusable knowledge. Default: `2048`.
+- `defaultModelContextTokens?: number` - used when `getHistoryWindow` receives no model size; a warning is logged. Default: `262144`.
+- `maxHistoryAgeMs?: number` - maximum Topic and raw-message age restored during lazy cold start. Default: `0` (all history).
+- `sessionIdleTtlMs?: number` - idle time before a session cache is released. Persistent data is not deleted. Default: `1800000`.
+- `sessionSweepIntervalMs?: number` - idle cache sweep interval. Default: `60000`.
+- `maxConcurrentCompressions?: number` - global semaphore limit for concurrent compression and knowledge ingestion tasks. Default: `3`.
+- `entitySimilarityThreshold?: number` - graph entity similarity threshold. Default: `0.92`.
+- `defaultSearchLimit?: number` - default `search` result limit. Default: `10`.
+
+`sessionTokenLimit`, `historyWindowTokenLimit`, `topicRatio`, `detailMaxTokens`, and `conciseMaxTokens` remain only as deprecated 0.3.x configuration/read compatibility fields. New code should not use them.
 
 Knowledge-base fields:
 
@@ -341,21 +367,30 @@ Knowledge-base fields:
 
 The package intentionally separates durable base writes from expensive graph construction.
 
-`updateChat(messages, opts)`:
-
-- Embeds and writes raw messages to LanceDB before resolving.
-- Creates or updates the session record.
-- Updates the in-memory session cache.
-- If `sessionTokenLimit` is exceeded, compression is triggered in the background.
-- Background compression may create topics/history and graph entities later.
+`updateChat(messages, opts)`:
+
+- Registers its write synchronously, so a caller may intentionally fire-and-forget it and a following `getHistoryWindow` will still wait for that write.
+- Calculates tokens over `content`, `parts`, `payload`, and `metadata`, embeds the searchable text, and upserts raw messages by stable `messageId` before resolving.
+- Creates or updates the session record.
+- Updates the in-memory session cache.
+- Once `getHistoryWindow` has supplied the current model size, reaching the precompression threshold starts background compression.
+- Topic storage and cache replacement finish before conversation graph extraction; graph work remains asynchronous.
 
-`flushChat(sessionId?, opts?)`:
+`flushChat(sessionId?, opts?)`:
 
-- Forces compression for the current cached messages even if `sessionTokenLimit` was not reached.
+- Forces compression for the current cached raw messages even if the dynamic precompression threshold was not reached.
 - `wait: true` waits for topic creation, cache clearing, and history-window rebuild.
 - `wait: true` does not wait for graph extraction/persistence.
 - `waitGraph: true` waits for graph extraction/persistence too.
-- When `waitGraph: true`, graph waiting is bounded by `graphBuildTimeoutMs` and may throw a timeout error.
+- When `waitGraph: true`, graph waiting is bounded by `graphBuildTimeoutMs` and may throw a timeout error.
+
+`getHistoryWindow(sessionId, modelContextTokens?)`:
+
+- Lazily hydrates the session from persistent Topics plus raw messages after the newest Topic boundary.
+- Waits for registered message writes. If the hard raw budget is exceeded, it also waits for or starts compression until the returned context fits.
+- Returns `{ compressedContext, recentMessages, usage }`. `recentMessages` preserves the `RawMessage` input shape, including `messageId`, `parts`, `payload`, `metadata`, and `createdAt`.
+- Topic budget is fixed by manager configuration; the raw-message budget is recalculated from the supplied model context size.
+- If accumulated Topics exceed their fixed budget, the oldest approximately half are summarized into one Topic. This Topic-to-Topic rollup does not rebuild graph data.
 
 `addDocument(opts)`:
 
@@ -381,8 +416,9 @@ Use `wait: true` when the next line of code must immediately call `searchKnowled
 - `sessionId?: string` - session id.
 - `type?: "text" | "image" | "file"` - default: `"text"`.
 - `content: string` - required plain text used for embedding and retrieval.
-- `parts?: ContentPart[]` - optional multimodal content parts; stored serialized.
-- `usage?: number` - token count; estimated with `tiktoken` if omitted.
+- `parts?: ContentPart[]` - optional multimodal content parts; stored serialized.
+- `payload?: unknown` - host-framework message payload stored and returned without interpretation.
+- `usage?: number` - token count; estimated with `tiktoken` if omitted.
 - `metadata?: Record<string, unknown>` - custom metadata stored as JSON.
 - `createdAt?: number` - Unix milliseconds; default is current time.
 
@@ -394,21 +430,23 @@ Use `wait: true` when the next line of code must immediately call `searchKnowled
 
 `StoredMessage` is persisted message output with required ids, serialized `parts`, serialized `metadata`, generated vector, and `createdAt`.
 
-`Topic` is a compressed memory item:
-
-- `title` - short topic title.
-- `detail` - rich summary.
-- `summary` - medium summary.
-- `concise` - short summary.
-- `startTime` and `endTime` - covered message time range.
-- `recallCount` - affects freshness ranking through `recallBoostMs`.
+`Topic` is a compressed memory item:
+
+- `title` - short topic title.
+- `summary` - the only compressed body; its actual length is importance-aware and bounded by `topicSummaryMaxTokens`.
+- `tokens` - summary token count stored when the Topic is created.
+- `startMessageId` / `endMessageId` - exact raw-message boundaries, in addition to timestamps.
+- `startTime` and `endTime` - covered message time range.
+- `recallCount` - incremented best-effort when search recalls the Topic; it is not used to reorder the continuous history window.
 
-`Fact` is a user/chat scoped manual memory:
+`Fact` is a user/chat scoped manual memory:
 
 - `level: "user" | "chat"`
 - `userId`, `chatId`, optional `sessionId`
-- `content`
-- `createdAt`
+- `content`
+- optional stable `key` for immediate cache + database upsert
+- `createdAt`
+- `updatedAt`
 
 `Entity` and `Relation` are graph records:
 
@@ -460,13 +498,13 @@ Scopes:
 - `scope: "user"` - filter by `userId`; pass `scopeId`.
 - `scope: "all"` - no domain filter; `scopeId` is ignored.
 
-Search modes:
+Search modes:
 
-- `mode: "fast"` - only LanceDB hybrid/vector retrieval. No graph search.
+- `mode: "fast"` - only vector-store hybrid/vector retrieval. No graph search.
 - `mode: "auto"` - conversation search asks the LLM whether graph search is needed; knowledge search triggers graph expansion only when top chunk score is high enough.
-- `mode: "all"` - include graph search and merge graph results with vector/full-text results.
-
-For deterministic low-latency calls, use `fast`. For richer entity relationship answers, use `auto` or `all`.
+- `mode: "all"` - include graph search and merge graph results with vector/full-text results.
+
+Conversation search always searches both compressed Topics and raw messages, including raw rows already represented by a Topic. For deterministic low-latency calls, use `fast`. For richer entity relationship answers, use `auto` or `all`.
 
 ## API Reference
 
@@ -474,9 +512,9 @@ For deterministic low-latency calls, use `fast`. For richer entity relationship
 
 Creates the manager and resolves defaults. It does not connect to storage until `init()`.
 
-### `init(): Promise<void>`
-
-Initializes tiktoken, LanceDB, Grafeo, facts cache, sessions cache, and restored history windows. Always call before using read/write APIs.
+### `init(): Promise<void>`
+
+Initializes tiktoken, the vector store (backend auto-detection happens here), Grafeo, and the facts cache. Conversation sessions are hydrated lazily on first access and evicted after the configured idle timeout. Always call before using read/write APIs.
 
 ### `updateChat(messages, opts?): Promise<void>`
 
@@ -569,13 +607,19 @@ Context:
 - `chatId?: string`
 - `userId?: string`
 
-### `getHistoryWindow(sessionId): string`
+### `getHistoryWindow(sessionId, modelContextTokens?): Promise<MemoryContextWindow>`
+
+Returns the complete model-history window:
+
+- `compressedContext: string` - chronological Topic summaries within the fixed compressed budget.
+- `recentMessages: MemoryRawMessage[]` - all currently uncompressed raw messages in chronological order, preserving the host payload and metadata.
+- `usage` - model size, usable history size, compressed usage, and dynamic raw-message budget/usage.
+
+Pass the active model's context size on every read. Omitting it uses `defaultModelContextTokens` (256K by default) and logs a warning.
 
-Returns the compressed history window for a session. This is prompt-ready text built from recent `Topic` records using `topicRatio` and `historyWindowTokenLimit`.
+### `getRecentMessages(sessionId, limit): Promise<StoredMessage[]>`
 
-### `getRecentMessages(sessionId, limit): Promise<StoredMessage[]>`
-
-Returns the globally latest `limit` messages for a session in chronological order (oldest to newest within the returned window). The storage query orders by `createdAt` and uses `messageId` as a deterministic same-timestamp tiebreak before applying the limit, so long sessions do not return an early candidate window.
+Returns the globally latest `limit` messages for a session in chronological order (oldest to newest within the returned window). The storage query orders by `createdAt` and uses `messageId` as a deterministic same-timestamp tiebreak before applying the limit, so long sessions do not return an early candidate window.
 
 ### Session APIs
 
@@ -590,7 +634,8 @@ Returns the globally latest `limit` messages for a session in chronological orde
 Session update options:
 
 - `title?: string`
-- `metadata?: Record<string, unknown>`
+- `metadata?: Record<string, unknown>`
+- `scope?: "session" | "chat" | "user" | "all"` - content-hash deduplication boundary; default is `session`.
 
 Session search options:
 
@@ -628,7 +673,7 @@ Options:
 
 Deduplication:
 
-- Documents are deduplicated by `sha256(content)` within the same `userId + chatId + sessionId` domain.
+- Documents are deduplicated by `sha256(content)` within the requested scope (`session` by default, or `chat` / `user` / `all`).
 - If a duplicate exists, `addDocument` returns the existing `docId`.
 
 Graph behavior:
@@ -673,11 +718,22 @@ Lists documents, optionally filtered by:
 - `chatId`
 - `sessionId`
 
-### `destroy(): void`
-
-Closes the Grafeo connection. It does not delete LanceDB or Grafeo data.
+### `destroy(): Promise<void>`
+
+Waits for registered writes, compression, graph work, and storage maintenance, then closes Grafeo and the vector store. It does not delete data.
 
-## Common Recipes
+## Common Recipes
+
+### Migrate A 0.3.x Store
+
+Stop all processes that write the store, then run the manual, idempotent migration. It backfills single-field Topic summaries/tokens/boundaries, raw-message usage, and fact timestamps without deleting raw messages or rerunning the LLM.
+
+```bash
+ppagent-memory migrate --config ./config.json --dry-run
+ppagent-memory migrate --config ./config.json --backup ./memory-backup
+```
+
+The config loader accepts a direct `MemoryConfig`, `memory`, `app.memory.auto`, or `service.memory` object. `--dry-run` migrates a temporary copy. `--backup` copies both supported vector-store files/directories plus Grafeo data and writes rollback guidance before mutating the live store.
 
 ### Store User And Assistant Messages
 
@@ -691,14 +747,20 @@ await memory.updateChat(
 );
 ```
 
-### Force A Prompt-Ready History Window
-
-```ts
-await memory.flushChat("s1", { wait: true });
-const history = memory.getHistoryWindow("s1");
-
-const systemPrompt = `你是助手。以下是长期记忆摘要：\n${history}`;
-```
+### Force A Prompt-Ready History Window
+
+```ts
+await memory.flushChat("s1", { wait: true });
+const history = await memory.getHistoryWindow("s1", 256 * 1024);
+
+const promptMessages = [
+  { role: "system", content: `更早对话摘要：\n${history.compressedContext}` },
+  ...history.recentMessages.map((message) => ({
+    role: message.talkerId === "assistant" ? "assistant" : "user",
+    content: message.payload ?? message.content,
+  })),
+];
+```
 
 ### Wait For Conversation Graph Data
 
@@ -806,35 +868,36 @@ Conversation graph search only targets conversation graph data. Knowledge graph
 ## Best Practices
 
 - Always call `await memory.init()` before using the manager.
-- Always call `memory.destroy()` during shutdown to close Grafeo resources.
+- Always call `await memory.destroy()` during shutdown to drain background work and close resources.
 - Use stable `userId`, `chatId`, and `sessionId` values. They define memory isolation and retrieval scope.
 - Use `scope: "user"` for personal memory, `scope: "chat"` for one agent/conversation channel, and `scope: "session"` for one short-lived conversation.
-- Use `getFactsForContext` for stable facts and `getHistoryWindow` for compressed session history.
+- Use `getFactsForContext` for stable facts and `getHistoryWindow` for compressed Topics plus recent raw session history.
 - Use `updateFacts` for explicit facts that should not depend on LLM extraction.
 - Use `updateEntity` when you already know structured entities and relations.
 - Use `flushChat(..., { wait: true })` before reading `getHistoryWindow` in tests or immediate workflows.
 - Use `addDocument(..., { wait: true })` before immediate knowledge search.
 - Use `waitGraph: true` sparingly because graph extraction depends on LLM calls and can be slower.
-- Keep `embeddingDimension` unchanged for an existing LanceDB directory unless you rebuild the database.
+- Keep `embeddingDimension` unchanged for an existing vector store unless you rebuild the database.
 - Keep examples free of real API keys, tokens, personal data, and proprietary customer content.
 
 ## Common Pitfalls
 
 - Forgetting `init()` before `updateChat` or search.
-- Expecting `updateChat` to finish compression or graph extraction. It only waits for raw message storage.
+- Treating `getHistoryWindow` as synchronous or string-only. Since 0.4 it is async and returns a structured complete window.
+- Omitting stable `messageId` when the host can update an existing assistant message; without it, an update becomes another raw message.
 - Expecting `flushChat(..., { wait: true })` to wait for graph data. Use `waitGraph: true`.
 - Expecting `addDocument(..., { wait: true })` to wait for graph data. Use `waitGraph: true`.
 - Setting `waitGraph: true` without allowing enough `graphBuildTimeoutMs` for large documents.
 - Using `scope: "session"` with a `chatId`, or `scope: "chat"` with a `userId`. `scopeId` must match the selected scope.
 - Searching immediately after `addDocument` without `wait: true`; chunks may still be ingesting.
-- Mixing embedding models with different vector dimensions in the same existing LanceDB path.
+- Mixing embedding models with different vector dimensions in the same existing vector store.
 - Treating `metadata` as indexed arbitrary JSON. Store important filter dimensions in stable top-level ids where possible.
 - Using `mode: "all"` for every query. It is richer but can be slower than `fast`.
 
 ## Security And Privacy
 
 - Never hard-code real `llmApiKey` or `embeddingApiKey`.
-- Treat LanceDB and Grafeo files as sensitive data stores.
+- Treat vector store (LanceDB/SQLite) and Grafeo files as sensitive data stores.
 - `content`, `metadata`, facts, and graph `meta` may contain private information.
 - `destroy()` closes resources; it does not scrub data.
 - Implement your own retention/deletion policy around `deleteSession`, `deleteDocument`, `deleteFact`, and storage directory cleanup.

package/package.json

@@ -1,10 +1,13 @@
 {
   "name": "@ppagent/memory",
-  "version": "0.1.4",
-  "description": "独立记忆系统模块，基于 LanceDB + Grafeo 实现向量存储与知识图谱",
+  "version": "0.4.0",
+  "description": "独立记忆系统模块，向量存储支持 LanceDB / SQLite(sqlite-vec) 双后端自动切换 + Grafeo 知识图谱",
   "type": "module",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
+  "bin": {
+    "ppagent-memory": "dist/cli.js"
+  },
   "exports": {
     ".": {
       "ppagent-source": "./src/index.ts",
@@ -25,29 +28,38 @@
     "memory",
     "rag",
     "lancedb",
+    "sqlite",
+    "sqlite-vec",
     "grafeo",
     "vector",
-    "graph"
+    "graph",
+    "hybrid-search"
   ],
   "author": "ppagent",
   "license": "Apache-2.0",
   "dependencies": {
     "@dqbd/tiktoken": "^1.0.22",
     "@grafeo-db/js": "^0.5.42",
-    "@lancedb/lancedb": "^0.30.0",
-    "apache-arrow": "^18.1.0",
+    "better-sqlite3": "^11.8.1",
+    "jieba-wasm": "^2.4.0",
+    "sqlite-vec": "^0.1.9",
     "uuid": "^10.0.0"
   },
   "devDependencies": {
+    "@types/better-sqlite3": "^7.6.13",
     "@types/node": "^22.0.0",
     "@types/uuid": "^10.0.0",
     "tsup": "^8.3.5",
     "typescript": "^5.8.3",
     "vitest": "^2.0.5"
   },
+  "optionalDependencies": {
+    "@lancedb/lancedb": "^0.30.0",
+    "apache-arrow": "^18.1.0"
+  },
   "scripts": {
-    "build": "tsup src/index.ts --format esm --dts --out-dir dist --clean",
-    "dev": "tsup src/index.ts --format esm --dts --out-dir dist --watch",
+    "build": "tsup src/index.ts --format esm --dts --out-dir dist --clean --external @lancedb/lancedb --external apache-arrow && tsup src/cli.ts --format esm --out-dir dist --external @lancedb/lancedb --external apache-arrow",
+    "dev": "tsup src/index.ts --format esm --dts --out-dir dist --watch --external @lancedb/lancedb --external apache-arrow",
     "test": "vitest"
   }
 }