A shared brain for AI agents that stays local

What it does

mcp-memory-service is a self-hosted memory backend for AI agent pipelines. Agents store decisions and facts via REST or MCP, then retrieve them in about five milliseconds across runs and frameworks like LangGraph, CrewAI, and AutoGen. Everything stays local: embeddings run via ONNX, storage is SQLite, and there is no per-call cloud rent.

The interesting bit

The service doubles as an ad-hoc message bus. Agents can tag memories with sentinels like msg:cluster, letting separate agents poll those tags to pick up cross-cluster signals—no custom protocol required. It also builds a knowledge graph with typed edges (causes, fixes, contradicts) and runs autonomous consolidation to compress old memories, which is more structure than typical vector-only stores.

Key highlights

• Framework-agnostic REST API with 76 endpoints, plus MCP transport for Claude Desktop and browser-based claude.ai via Remote MCP.
• Knowledge graph with typed edges and hybrid search (BM25 + vector) over turn-level or session-level storage.
• Local ONNX embeddings and SQLite backend; no cloud lock-in or API costs.
• Real-time SSE events notify clients when any agent stores or deletes a memory.
• X-Agent-ID headers auto-scope memories to specific agents for filtered retrieval.

Verdict

Worth a look if you’re running multi-agent pipelines that need shared state without managed cloud fees. Skip it if you already have a Redis-Pinecone stack you’re happy maintaining.