feat(Phase 4): Memory & Intelligence — memory, log, summarization, skills (#91)

Co-authored-by: Jason Woltje <jason@diversecanvas.com>
Co-committed-by: Jason Woltje <jason@diversecanvas.com>

packages/memory/package.json

@@ -1,6 +1,7 @@
 {
   "name": "@mosaic/memory",
   "version": "0.0.0",
+  "type": "module",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
   "exports": {
@@ -16,7 +17,9 @@
     "test": "vitest run --passWithNoTests"
   },
   "dependencies": {
-    "@mosaic/types": "workspace:*"
+    "@mosaic/db": "workspace:*",
+    "@mosaic/types": "workspace:*",
+    "drizzle-orm": "^0.45.1"
   },
   "devDependencies": {
     "typescript": "^5.8.0",

packages/memory/src/index.ts

@@ -1 +1,15 @@
-export const VERSION = '0.0.0';
+export { createMemory, type Memory } from './memory.js';
+export {
+  createPreferencesRepo,
+  type PreferencesRepo,
+  type Preference,
+  type NewPreference,
+} from './preferences.js';
+export {
+  createInsightsRepo,
+  type InsightsRepo,
+  type Insight,
+  type NewInsight,
+  type SearchResult,
+} from './insights.js';
+export type { VectorStore, VectorSearchResult, EmbeddingProvider } from './vector-store.js';

packages/memory/src/insights.ts

@@ -0,0 +1,89 @@
+import { eq, and, desc, sql, lt, type Db, insights } from '@mosaic/db';
+
+export type Insight = typeof insights.$inferSelect;
+export type NewInsight = typeof insights.$inferInsert;
+
+export interface SearchResult {
+  insight: Insight;
+  distance: number;
+}
+
+export function createInsightsRepo(db: Db) {
+  return {
+    async findByUser(userId: string, limit = 50): Promise<Insight[]> {
+      return db
+        .select()
+        .from(insights)
+        .where(eq(insights.userId, userId))
+        .orderBy(desc(insights.createdAt))
+        .limit(limit);
+    },
+
+    async findById(id: string): Promise<Insight | undefined> {
+      const rows = await db.select().from(insights).where(eq(insights.id, id));
+      return rows[0];
+    },
+
+    async create(data: NewInsight): Promise<Insight> {
+      const rows = await db.insert(insights).values(data).returning();
+      return rows[0]!;
+    },
+
+    async update(id: string, data: Partial<NewInsight>): Promise<Insight | undefined> {
+      const rows = await db
+        .update(insights)
+        .set({ ...data, updatedAt: new Date() })
+        .where(eq(insights.id, id))
+        .returning();
+      return rows[0];
+    },
+
+    async remove(id: string): Promise<boolean> {
+      const rows = await db.delete(insights).where(eq(insights.id, id)).returning();
+      return rows.length > 0;
+    },
+
+    /**
+     * Semantic search using pgvector cosine distance.
+     * Requires the vector extension and an embedding for the query.
+     */
+    async searchByEmbedding(
+      userId: string,
+      queryEmbedding: number[],
+      limit = 10,
+      maxDistance = 0.8,
+    ): Promise<SearchResult[]> {
+      const embeddingStr = `[${queryEmbedding.join(',')}]`;
+      const rows = await db.execute(sql`
+        SELECT *,
+               (embedding <=> ${embeddingStr}::vector) AS distance
+        FROM insights
+        WHERE user_id = ${userId}
+          AND embedding IS NOT NULL
+          AND (embedding <=> ${embeddingStr}::vector) < ${maxDistance}
+        ORDER BY distance ASC
+        LIMIT ${limit}
+      `);
+
+      return rows as unknown as SearchResult[];
+    },
+
+    /**
+     * Decay relevance scores for old insights that haven't been accessed recently.
+     */
+    async decayOldInsights(olderThan: Date, decayFactor = 0.95): Promise<number> {
+      const result = await db
+        .update(insights)
+        .set({
+          relevanceScore: sql`${insights.relevanceScore} * ${decayFactor}`,
+          decayedAt: new Date(),
+          updatedAt: new Date(),
+        })
+        .where(and(lt(insights.updatedAt, olderThan), sql`${insights.relevanceScore} > 0.1`))
+        .returning();
+      return result.length;
+    },
+  };
+}
+
+export type InsightsRepo = ReturnType<typeof createInsightsRepo>;

packages/memory/src/memory.ts

@@ -0,0 +1,15 @@
+import type { Db } from '@mosaic/db';
+import { createPreferencesRepo, type PreferencesRepo } from './preferences.js';
+import { createInsightsRepo, type InsightsRepo } from './insights.js';
+
+export interface Memory {
+  preferences: PreferencesRepo;
+  insights: InsightsRepo;
+}
+
+export function createMemory(db: Db): Memory {
+  return {
+    preferences: createPreferencesRepo(db),
+    insights: createInsightsRepo(db),
+  };
+}

packages/memory/src/preferences.ts

@@ -0,0 +1,59 @@
+import { eq, and, type Db, preferences } from '@mosaic/db';
+
+export type Preference = typeof preferences.$inferSelect;
+export type NewPreference = typeof preferences.$inferInsert;
+
+export function createPreferencesRepo(db: Db) {
+  return {
+    async findByUser(userId: string): Promise<Preference[]> {
+      return db.select().from(preferences).where(eq(preferences.userId, userId));
+    },
+
+    async findByUserAndKey(userId: string, key: string): Promise<Preference | undefined> {
+      const rows = await db
+        .select()
+        .from(preferences)
+        .where(and(eq(preferences.userId, userId), eq(preferences.key, key)));
+      return rows[0];
+    },
+
+    async findByUserAndCategory(
+      userId: string,
+      category: Preference['category'],
+    ): Promise<Preference[]> {
+      return db
+        .select()
+        .from(preferences)
+        .where(and(eq(preferences.userId, userId), eq(preferences.category, category)));
+    },
+
+    async upsert(data: NewPreference): Promise<Preference> {
+      const existing = await db
+        .select()
+        .from(preferences)
+        .where(and(eq(preferences.userId, data.userId), eq(preferences.key, data.key)));
+
+      if (existing[0]) {
+        const rows = await db
+          .update(preferences)
+          .set({ value: data.value, category: data.category, updatedAt: new Date() })
+          .where(eq(preferences.id, existing[0].id))
+          .returning();
+        return rows[0]!;
+      }
+
+      const rows = await db.insert(preferences).values(data).returning();
+      return rows[0]!;
+    },
+
+    async remove(userId: string, key: string): Promise<boolean> {
+      const rows = await db
+        .delete(preferences)
+        .where(and(eq(preferences.userId, userId), eq(preferences.key, key)))
+        .returning();
+      return rows.length > 0;
+    },
+  };
+}
+
+export type PreferencesRepo = ReturnType<typeof createPreferencesRepo>;

packages/memory/src/vector-store.ts

@@ -0,0 +1,39 @@
+/**
+ * VectorStore interface — abstraction over pgvector that allows future
+ * swap to Qdrant, Pinecone, etc.
+ */
+export interface VectorStore {
+  /** Store an embedding with an associated document ID. */
+  store(documentId: string, embedding: number[], metadata?: Record<string, unknown>): Promise<void>;
+
+  /** Search for similar embeddings, returning document IDs and distances. */
+  search(
+    queryEmbedding: number[],
+    limit?: number,
+    filter?: Record<string, unknown>,
+  ): Promise<VectorSearchResult[]>;
+
+  /** Delete an embedding by document ID. */
+  remove(documentId: string): Promise<void>;
+}
+
+export interface VectorSearchResult {
+  documentId: string;
+  distance: number;
+  metadata?: Record<string, unknown>;
+}
+
+/**
+ * EmbeddingProvider interface — generates embeddings from text.
+ * Implemented by the gateway using the configured LLM provider.
+ */
+export interface EmbeddingProvider {
+  /** Generate an embedding vector for the given text. */
+  embed(text: string): Promise<number[]>;
+
+  /** Generate embeddings for multiple texts in batch. */
+  embedBatch(texts: string[]): Promise<number[][]>;
+
+  /** The dimensionality of the embeddings this provider generates. */
+  dimensions: number;
+}