@jungjaehoon/mama-core 1.1.5 → 1.2.1

package/README.md

@@ -108,7 +108,7 @@ const mamaApi = require('@jungjaehoon/mama-core/mama-api');
 
 - **@huggingface/transformers** - Local embedding generation
 - **better-sqlite3** - SQLite database
-- **sqlite-vec** - Vector similarity extension
+- **Pure-TS cosine similarity** - Vector search (no native extensions)
 
 ## Development
 

package/db/migrations/013-replace-vss-with-embeddings.sql

@@ -0,0 +1,12 @@
+-- Migration 013: Replace vss_memories (sqlite-vec virtual table) with plain embeddings table
+-- Vector search now uses pure TypeScript brute-force cosine similarity.
+-- No native extensions required.
+
+CREATE TABLE IF NOT EXISTS embeddings (
+  rowid INTEGER PRIMARY KEY,
+  embedding BLOB NOT NULL
+);
+
+-- Note: Data migration from vss_memories is handled programmatically in sqlite-adapter.ts
+-- because vss_memories is a vec0 virtual table that requires sqlite-vec to read.
+-- If sqlite-vec is available at migration time, the adapter will copy rows automatically.

package/dist/config-loader.d.ts

@@ -15,6 +15,7 @@
 export interface MAMAConfig {
     modelName: string;
     embeddingDim: number;
+    quantized: boolean;
     cacheDir: string;
     [key: string]: unknown;
 }
@@ -36,6 +37,10 @@ export declare function getModelName(): string;
  * Get current embedding dimension
  */
 export declare function getEmbeddingDim(): number;
+/**
+ * Get whether to use quantized model
+ */
+export declare function getQuantized(): boolean;
 /**
  * Get current cache directory
  */
@@ -43,6 +48,7 @@ export declare function getCacheDir(): string;
 export interface ConfigUpdates {
     modelName?: string;
     embeddingDim?: number;
+    quantized?: boolean;
     cacheDir?: string;
 }
 /**

package/dist/config-loader.js

@@ -21,6 +21,7 @@ exports.DEFAULT_CONFIG = void 0;
 exports.loadConfig = loadConfig;
 exports.getModelName = getModelName;
 exports.getEmbeddingDim = getEmbeddingDim;
+exports.getQuantized = getQuantized;
 exports.getCacheDir = getCacheDir;
 exports.updateConfig = updateConfig;
 exports.getConfigPath = getConfigPath;
@@ -32,6 +33,7 @@ const debug_logger_js_1 = require("./debug-logger.js");
 exports.DEFAULT_CONFIG = {
     modelName: 'Xenova/multilingual-e5-small',
     embeddingDim: 384,
+    quantized: true,
     cacheDir: path_1.default.join(os_1.default.homedir(), '.cache', 'huggingface', 'transformers'),
 };
 // Config file path
@@ -129,6 +131,13 @@ function getEmbeddingDim() {
     const config = loadConfig();
     return config.embeddingDim;
 }
+/**
+ * Get whether to use quantized model
+ */
+function getQuantized() {
+    const config = loadConfig();
+    return config.quantized;
+}
 /**
  * Get current cache directory
  */

package/dist/db-adapter/sqlite-adapter.d.ts

@@ -1,8 +1,8 @@
 /**
  * SQLite Database Adapter
  *
- * Implements DatabaseAdapter interface using better-sqlite3 + sqlite-vec
- * This is the current production implementation extracted from memory-store.js
+ * Implements DatabaseAdapter interface using better-sqlite3
+ * Vector search uses pure TypeScript brute-force cosine similarity (no native extensions)
  *
  * @module sqlite-adapter
  */
@@ -53,11 +53,13 @@ export declare class SQLiteAdapter extends DatabaseAdapter {
      */
     transaction<T>(fn: () => T): T;
     /**
-     * Vector similarity search using sqlite-vec (vec0 virtual table)
+     * Vector similarity search using brute-force cosine similarity
+     * Loads all embeddings from the embeddings table, computes similarity in JS,
+     * and returns top-N results sorted by similarity descending.
      */
     vectorSearch(embedding: Float32Array | number[], limit?: number): VectorSearchResult[] | null;
     /**
-     * Insert vector embedding
+     * Insert vector embedding into the embeddings table (plain BLOB storage)
      */
     insertEmbedding(rowid: number, embedding: Float32Array | number[]): RunResult | null;
     /**
@@ -68,6 +70,11 @@ export declare class SQLiteAdapter extends DatabaseAdapter {
      * Run migrations
      */
     runMigrations(migrationsDir: string): void;
+    /**
+     * Attempt to migrate data from legacy vss_memories (sqlite-vec) to embeddings table.
+     * This is best-effort: if sqlite-vec is not loaded, the vss_memories table won't be readable.
+     */
+    private _migrateFromVssMemories;
 }
 export default SQLiteAdapter;
 //# sourceMappingURL=sqlite-adapter.d.ts.map

package/dist/db-adapter/sqlite-adapter.js

@@ -2,8 +2,8 @@
 /**
  * SQLite Database Adapter
  *
- * Implements DatabaseAdapter interface using better-sqlite3 + sqlite-vec
- * This is the current production implementation extracted from memory-store.js
+ * Implements DatabaseAdapter interface using better-sqlite3
+ * Vector search uses pure TypeScript brute-force cosine similarity (no native extensions)
  *
  * @module sqlite-adapter
  */
@@ -20,16 +20,6 @@ const base_adapter_js_1 = require("./base-adapter.js");
 const statement_js_1 = require("./statement.js");
 const debug_logger_js_1 = require("../debug-logger.js");
 const embeddings_js_1 = require("../embeddings.js");
-// Try to load sqlite-vec
-let sqliteVec = null;
-try {
-    // eslint-disable-next-line @typescript-eslint/no-require-imports
-    sqliteVec = require('sqlite-vec');
-}
-catch {
-    // Defer logging until connect() so we have logger context initialized
-    sqliteVec = null;
-}
 // Database paths
 const LEGACY_DB_PATH = path_1.default.join(os_1.default.homedir(), '.spinelift', 'memories.db');
 // Default to ~/.claude/mama-memory.db for Claude Code/Desktop compatibility
@@ -37,7 +27,7 @@ const DEFAULT_DB_PATH = path_1.default.join(os_1.default.homedir(), '.claude', '
 class SQLiteAdapter extends base_adapter_js_1.DatabaseAdapter {
     config;
     db = null;
-    _vectorSearchEnabled = false;
+    _vectorSearchEnabled = true; // Always true: pure TS cosine similarity
     constructor(config = {}) {
         super();
         this.config = config;
@@ -93,23 +83,7 @@ class SQLiteAdapter extends base_adapter_js_1.DatabaseAdapter {
         this.db.pragma('cache_size = -64000'); // 64MB cache
         this.db.pragma('temp_store = MEMORY');
         this.db.pragma('foreign_keys = ON');
-        // Load sqlite-vec extension (graceful degradation if unavailable)
-        if (sqliteVec) {
-            try {
-                sqliteVec.load(this.db);
-                this._vectorSearchEnabled = true;
-                (0, debug_logger_js_1.info)('[sqlite-adapter] Loaded sqlite-vec extension');
-            }
-            catch (err) {
-                this._vectorSearchEnabled = false;
-                const message = err instanceof Error ? err.message : String(err);
-                (0, debug_logger_js_1.warn)(`[sqlite-adapter] sqlite-vec unavailable (Tier 2 fallback): ${message}`);
-            }
-        }
-        else {
-            this._vectorSearchEnabled = false;
-            (0, debug_logger_js_1.warn)('[sqlite-adapter] sqlite-vec package not installed; vector search disabled');
-        }
+        (0, debug_logger_js_1.info)('[sqlite-adapter] Vector search: pure TS cosine similarity (no native extensions)');
         return this.db;
     }
     /**
@@ -158,64 +132,55 @@ class SQLiteAdapter extends base_adapter_js_1.DatabaseAdapter {
         return txn();
     }
     /**
-     * Vector similarity search using sqlite-vec (vec0 virtual table)
+     * Vector similarity search using brute-force cosine similarity
+     * Loads all embeddings from the embeddings table, computes similarity in JS,
+     * and returns top-N results sorted by similarity descending.
      */
     vectorSearch(embedding, limit = 5) {
         if (!this.isConnected()) {
             throw new Error('Database not connected');
         }
-        if (!this._vectorSearchEnabled) {
-            return null;
-        }
-        const embeddingJson = JSON.stringify(Array.from(embedding));
-        const stmt = this.prepare(`
-      SELECT
-        rowid,
-        embedding,
-        distance
-      FROM vss_memories
-      WHERE embedding MATCH vec_f32(?)
-      LIMIT ?
-    `);
         const queryVector = embedding instanceof Float32Array ? embedding : Float32Array.from(embedding);
-        const results = stmt.all(embeddingJson, Math.max(limit, 1));
-        return results
-            .map((row) => {
+        // Check if embeddings table exists
+        const tableCheck = this.prepare(`SELECT name FROM sqlite_master WHERE type='table' AND name='embeddings'`).all();
+        if (tableCheck.length === 0) {
+            throw new Error('Embeddings table missing');
+        }
+        const rows = this.prepare('SELECT rowid, embedding FROM embeddings').all();
+        const scored = [];
+        for (const row of rows) {
             const candidate = bufferToVector(row.embedding);
-            if (!candidate) {
-                return null;
+            if (!candidate)
+                continue;
+            if (candidate.length !== queryVector.length) {
+                (0, debug_logger_js_1.warn)(`Skipping rowid ${row.rowid}: dimension mismatch (${candidate.length} vs ${queryVector.length})`);
+                continue;
             }
             const similarity = (0, embeddings_js_1.cosineSimilarity)(candidate, queryVector);
-            return {
+            scored.push({
                 rowid: row.rowid,
                 similarity,
                 distance: 1 - similarity,
-            };
-        })
-            .filter((r) => r !== null);
+            });
+        }
+        // Sort by similarity descending, return top-N
+        scored.sort((a, b) => (b.similarity ?? 0) - (a.similarity ?? 0));
+        return scored.slice(0, Math.max(limit, 1));
     }
     /**
-     * Insert vector embedding
+     * Insert vector embedding into the embeddings table (plain BLOB storage)
      */
     insertEmbedding(rowid, embedding) {
         if (!this.isConnected()) {
             throw new Error('Database not connected');
         }
-        if (!this._vectorSearchEnabled) {
-            return null;
-        }
-        const embeddingJson = JSON.stringify(Array.from(embedding));
-        // CRITICAL FIX: sqlite-vec virtual tables accept rowid as literal but not via ? placeholder
-        // Using template literal with Number() cast for safety (prevents SQL injection)
-        const safeRowid = Number(rowid);
-        if (!Number.isInteger(safeRowid) || safeRowid < 1) {
-            throw new Error(`Invalid rowid: ${rowid}`);
-        }
+        const vec = embedding instanceof Float32Array ? embedding : Float32Array.from(embedding);
+        const buffer = Buffer.from(vec.buffer, vec.byteOffset, vec.byteLength);
         const stmt = this.prepare(`
-      INSERT OR REPLACE INTO vss_memories(rowid, embedding)
-      VALUES (${safeRowid}, ?)
+      INSERT OR REPLACE INTO embeddings(rowid, embedding)
+      VALUES (?, ?)
     `);
-        return stmt.run(embeddingJson);
+        return stmt.run(rowid, buffer);
     }
     /**
      * Get last inserted row ID
@@ -298,33 +263,61 @@ class SQLiteAdapter extends base_adapter_js_1.DatabaseAdapter {
                 throw new Error(`Migration ${file} failed: ${message}`);
             }
         }
-        // Create vss_memories table if not exists
-        if (this._vectorSearchEnabled) {
-            const vssTables = this.prepare(`
-        SELECT name FROM sqlite_master
-        WHERE type='table' AND name='vss_memories'
-      `).all();
+        // Ensure embeddings table exists (plain table, no native extensions)
+        const embeddingsTables = this.prepare(`SELECT name FROM sqlite_master WHERE type='table' AND name='embeddings'`).all();
+        if (embeddingsTables.length === 0) {
+            (0, debug_logger_js_1.info)('[sqlite-adapter] Creating embeddings table');
+            this.exec(`
+        CREATE TABLE embeddings (
+          rowid INTEGER PRIMARY KEY,
+          embedding BLOB NOT NULL
+        )
+      `);
+        }
+        // Always attempt migration — uses INSERT OR IGNORE so safe to run multiple times
+        this._migrateFromVssMemories();
+    }
+    /**
+     * Attempt to migrate data from legacy vss_memories (sqlite-vec) to embeddings table.
+     * This is best-effort: if sqlite-vec is not loaded, the vss_memories table won't be readable.
+     */
+    _migrateFromVssMemories() {
+        try {
+            const vssTables = this.prepare(`SELECT name FROM sqlite_master WHERE name='vss_memories'`).all();
             if (vssTables.length === 0) {
-                (0, debug_logger_js_1.info)('[sqlite-adapter] Creating vss_memories virtual table via sqlite-vec');
-                this.exec(`
-          CREATE VIRTUAL TABLE vss_memories USING vec0(
-            embedding float[384]
-          )
-        `);
+                return; // No legacy table
+            }
+            // Try to read from vss_memories — will fail if sqlite-vec extension isn't loaded
+            const rows = this.prepare('SELECT rowid, embedding FROM vss_memories').all();
+            if (rows.length > 0) {
+                const insertStmt = this.prepare('INSERT OR IGNORE INTO embeddings (rowid, embedding) VALUES (?, ?)');
+                let migrated = 0;
+                for (const row of rows) {
+                    const res = insertStmt.run(row.rowid, row.embedding);
+                    if (res.changes > 0)
+                        migrated++;
+                }
+                (0, debug_logger_js_1.info)(`[sqlite-adapter] Migrated ${migrated} embeddings from vss_memories to embeddings`);
             }
         }
-        else {
-            (0, debug_logger_js_1.warn)('[sqlite-adapter] Skipping vss_memories creation (sqlite-vec unavailable)');
+        catch (err) {
+            const message = err instanceof Error ? err.message : String(err);
+            (0, debug_logger_js_1.warn)(`[sqlite-adapter] Could not migrate from vss_memories (expected if sqlite-vec not installed): ${message}`);
         }
     }
 }
 exports.SQLiteAdapter = SQLiteAdapter;
 exports.default = SQLiteAdapter;
 function bufferToVector(buffer) {
-    if (!buffer) {
+    if (!buffer || buffer.byteLength % 4 !== 0) {
+        return null;
+    }
+    try {
+        const arrayBuffer = buffer.buffer.slice(buffer.byteOffset, buffer.byteOffset + buffer.byteLength);
+        return new Float32Array(arrayBuffer);
+    }
+    catch {
         return null;
     }
-    const arrayBuffer = buffer.buffer.slice(buffer.byteOffset, buffer.byteOffset + buffer.byteLength);
-    return new Float32Array(arrayBuffer);
 }
 //# sourceMappingURL=sqlite-adapter.js.map

package/dist/db-manager.d.ts

@@ -2,7 +2,7 @@
  * MAMA Database Manager (SQLite-only)
  *
  * SQLite-exclusive database interface for MAMA Plugin.
- * Uses better-sqlite3 + sqlite-vec for local storage.
+ * Uses better-sqlite3 for local storage.
  *
  * PostgreSQL support is only available in the legacy mcp-server repository.
  *
@@ -10,7 +10,7 @@
  * - WAL mode for better concurrency
  * - synchronous=NORMAL for performance
  * - Automatic migration management
- * - Vector similarity search (when sqlite-vec available)
+ * - Vector similarity search (pure TS cosine similarity)
  *
  * @module db-manager
  * @version 2.1 (Plugin - SQLite-only)
@@ -143,8 +143,7 @@ export declare function closeDB(): Promise<void>;
 /**
  * Insert embedding into vector search table
  *
- * Uses sqlite-vec for vector similarity search
- * Gracefully degrades if sqlite-vec is not available
+ * Stores embedding in the embeddings table for vector similarity search
  *
  * @param decisionRowid - SQLite rowid
  * @param embedding - 384-dim embedding vector

package/dist/db-manager.js

@@ -3,7 +3,7 @@
  * MAMA Database Manager (SQLite-only)
  *
  * SQLite-exclusive database interface for MAMA Plugin.
- * Uses better-sqlite3 + sqlite-vec for local storage.
+ * Uses better-sqlite3 for local storage.
  *
  * PostgreSQL support is only available in the legacy mcp-server repository.
  *
@@ -11,7 +11,7 @@
  * - WAL mode for better concurrency
  * - synchronous=NORMAL for performance
  * - Automatic migration management
- * - Vector similarity search (when sqlite-vec available)
+ * - Vector similarity search (pure TS cosine similarity)
  *
  * @module db-manager
  * @version 2.1 (Plugin - SQLite-only)
@@ -142,8 +142,7 @@ async function closeDB() {
 /**
  * Insert embedding into vector search table
  *
- * Uses sqlite-vec for vector similarity search
- * Gracefully degrades if sqlite-vec is not available
+ * Stores embedding in the embeddings table for vector similarity search
  *
  * @param decisionRowid - SQLite rowid
  * @param embedding - 384-dim embedding vector
@@ -172,7 +171,7 @@ async function insertEmbedding(decisionRowid, embedding) {
 async function vectorSearch(queryEmbedding, limit = 5, threshold = 0.7) {
     const adapter = getAdapter();
     try {
-        // SQLite adapter returns null if sqlite-vec not available
+        // Brute-force cosine similarity over all embeddings
         const results = await adapter.vectorSearch(queryEmbedding, limit * 3);
         if (!results || results.length === 0) {
             return []; // No keyword fallback - fast fail

package/dist/decision-tracker.d.ts

@@ -185,7 +185,7 @@ export declare function getSupersededChainDepth(topic: string): Promise<{
  * Task 3.5: If previous exists: Create supersedes edge, Update previous superseded_by
  * Task 3.6: If multi-parent refinement: Store refined_from, Calculate combined confidence
  * Task 3.7: Generate enhanced embedding
- * Task 3.8: Store in vss_memories (link via rowid)
+ * Task 3.8: Store embedding (link via rowid)
  *
  * AC #1: Decision stored with outcome=NULL, confidence from LLM
  * AC #2: Supersedes relationship creation

package/dist/decision-tracker.js

@@ -380,7 +380,7 @@ async function getSupersededChainDepth(topic) {
  * Task 3.5: If previous exists: Create supersedes edge, Update previous superseded_by
  * Task 3.6: If multi-parent refinement: Store refined_from, Calculate combined confidence
  * Task 3.7: Generate enhanced embedding
- * Task 3.8: Store in vss_memories (link via rowid)
+ * Task 3.8: Store embedding (link via rowid)
  *
  * AC #1: Decision stored with outcome=NULL, confidence from LLM
  * AC #2: Supersedes relationship creation
@@ -461,7 +461,7 @@ async function learnDecision(detection, toolExecution, sessionContext) {
                 : null,
             risks: detection.risks || null,
         };
-        // Task 3.7, 3.8: Generate enhanced embedding and store in vss_memories
+        // Task 3.7, 3.8: Generate enhanced embedding and store in embeddings table
         // (Handled by insertDecisionWithEmbedding function)
         await (0, memory_store_js_1.insertDecisionWithEmbedding)(decision);
         // ════════════════════════════════════════════════════════

package/dist/embeddings.d.ts

@@ -10,7 +10,7 @@
  * @date 2025-11-20
  */
 import { embeddingCache } from './embedding-cache.js';
-import { loadConfig, getModelName, getEmbeddingDim } from './config-loader.js';
+import { loadConfig, getModelName, getEmbeddingDim, getQuantized } from './config-loader.js';
 /**
  * Decision object for enhanced embedding generation
  */
@@ -71,5 +71,5 @@ export declare function cosineSimilarity(embA: Float32Array, embB: Float32Array)
 export { embeddingCache };
 export declare const EMBEDDING_DIM: number;
 export declare const MODEL_NAME: string;
-export { loadConfig, getModelName, getEmbeddingDim };
+export { loadConfig, getModelName, getEmbeddingDim, getQuantized };
 //# sourceMappingURL=embeddings.d.ts.map

package/dist/embeddings.js

@@ -14,7 +14,7 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
     return (mod && mod.__esModule) ? mod : { "default": mod };
 };
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.getEmbeddingDim = exports.getModelName = exports.loadConfig = exports.MODEL_NAME = exports.EMBEDDING_DIM = exports.embeddingCache = void 0;
+exports.getQuantized = exports.getEmbeddingDim = exports.getModelName = exports.loadConfig = exports.MODEL_NAME = exports.EMBEDDING_DIM = exports.embeddingCache = void 0;
 exports.generateEmbedding = generateEmbedding;
 exports.generateEnhancedEmbedding = generateEnhancedEmbedding;
 exports.generateBatchEmbeddings = generateBatchEmbeddings;
@@ -29,6 +29,7 @@ const config_loader_js_1 = require("./config-loader.js");
 Object.defineProperty(exports, "loadConfig", { enumerable: true, get: function () { return config_loader_js_1.loadConfig; } });
 Object.defineProperty(exports, "getModelName", { enumerable: true, get: function () { return config_loader_js_1.getModelName; } });
 Object.defineProperty(exports, "getEmbeddingDim", { enumerable: true, get: function () { return config_loader_js_1.getEmbeddingDim; } });
+Object.defineProperty(exports, "getQuantized", { enumerable: true, get: function () { return config_loader_js_1.getQuantized; } });
 // Shared cache directory (not in node_modules)
 const DEFAULT_CACHE_DIR = path_1.default.join(os_1.default.homedir(), '.cache', 'huggingface', 'transformers');
 // Singleton pattern for model loading
@@ -73,11 +74,14 @@ async function loadModel() {
             const cacheDir = process.env.HF_HOME || process.env.TRANSFORMERS_CACHE || DEFAULT_CACHE_DIR;
             env.cacheDir = cacheDir;
             (0, debug_logger_js_1.info)(`[MAMA] Model cache directory: ${cacheDir}`);
-            embeddingPipeline = (await pipeline('feature-extraction', modelName));
+            const quantized = (0, config_loader_js_1.getQuantized)();
+            embeddingPipeline = (await pipeline('feature-extraction', modelName, {
+                dtype: quantized ? 'q8' : 'fp32',
+            }));
             currentModelName = modelName;
             const loadTime = Date.now() - startTime;
             const config = (0, config_loader_js_1.loadConfig)();
-            (0, progress_indicator_js_1.logComplete)(`Embedding model ready (${loadTime}ms, ${config.embeddingDim}-dim)`);
+            (0, progress_indicator_js_1.logComplete)(`Embedding model ready (${loadTime}ms, ${config.embeddingDim}-dim, ${quantized ? 'q8' : 'fp32'})`);
         }
         catch (loadErr) {
             modelLoadFailed = true;

package/dist/mama-api.js

@@ -681,7 +681,7 @@ async function suggest(userQuestion, options = {}) {
         let results = [];
         let searchMethod = 'vector';
         try {
-            // Check if vector search is available (sqlite-vec loaded)
+            // Check if vector search is available
             if (!(0, memory_store_js_1.getAdapter)().vectorSearchEnabled) {
                 throw new Error('Vector search not available');
             }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@jungjaehoon/mama-core",
-  "version": "1.1.5",
+  "version": "1.2.1",
   "description": "MAMA Core - Shared modules for Memory-Augmented MCP Assistant",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -62,11 +62,8 @@
     "node": ">=18.0.0"
   },
   "dependencies": {
-    "@anthropic-ai/sdk": "^0.72.1",
     "@huggingface/transformers": "^3.8.1",
     "better-sqlite3": "^11.0.0",
-    "mime-types": "^3.0.2",
-    "sqlite-vec": "^0.1.0",
     "ws": "^8.19.0"
   },
   "devDependencies": {