feat: Replace JSON memory with pgvector semantic search (MAT-11)

Add memory-service (FastAPI + pgvector) for semantic memory storage.
Bot now queries relevant memories per conversation instead of dumping all 50.
Includes migration script for existing JSON files.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

docker-compose.yml

@@ -17,8 +17,45 @@ services:
       - DEFAULT_MODEL
       - WILDFILES_BASE_URL
       - WILDFILES_ORG
+      - MEMORY_SERVICE_URL=http://memory-service:8090
     volumes:
       - bot-data:/data
+    depends_on:
+      memory-service:
+        condition: service_healthy
+
+  memory-db:
+    image: pgvector/pgvector:pg17
+    restart: unless-stopped
+    environment:
+      POSTGRES_USER: memory
+      POSTGRES_PASSWORD: ${MEMORY_DB_PASSWORD:-memory}
+      POSTGRES_DB: memories
+    volumes:
+      - memory-pgdata:/var/lib/postgresql/data
+    healthcheck:
+      test: ["CMD-SHELL", "pg_isready -U memory -d memories"]
+      interval: 5s
+      timeout: 3s
+      retries: 5
+
+  memory-service:
+    build: ./memory-service
+    restart: unless-stopped
+    environment:
+      DATABASE_URL: postgresql://memory:${MEMORY_DB_PASSWORD:-memory}@memory-db:5432/memories
+      LITELLM_BASE_URL: ${LITELLM_BASE_URL}
+      LITELLM_API_KEY: ${LITELLM_API_KEY:-not-needed}
+      EMBED_MODEL: ${EMBED_MODEL:-text-embedding-3-small}
+    depends_on:
+      memory-db:
+        condition: service_healthy
+    healthcheck:
+      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8090/health')"]
+      interval: 10s
+      timeout: 5s
+      retries: 3
 
 volumes:
   bot-data:
+  memory-pgdata: