Skip to main content
Back

Open-Source mcp-memory-service Gives AI Agents 5ms Persistent Memory Without Cloud Lock-in

A self-hosted memory backend gives LangGraph, CrewAI and Claude agents persistent state, typed knowledge graphs, and 5ms recall with zero cloud cost.

Repo, Full Access
  • Open-source persistent memory backend for AI agents under Apache 2.0, works with LangGraph, CrewAI, AutoGen and Claude
  • Install via pip install mcp-memory-service, exposes both MCP and a 76-endpoint REST API
  • Local ONNX embeddings, SQLite-vec by default, optional Cloudflare or Milvus backends for teams
  • Knowledge graph with typed edges, autonomous consolidation, and 3D interactive visualization dashboard
  • Retrieval benchmarks: 80.4% R@5 on LongMemEval, 91.1% R@5 on DevBench, weaker on temporal queries
  • Latest release v11.8.0 fixes entity graph population and adds TLS verification by default

Every fresh chat session is a lobotomy. Your agent forgets the architecture, the rate limits, the bug you patched yesterday, and you spend the first ten minutes re-explaining context. mcp-memory-service is an open-source memory backend that tries to fix this at the infrastructure layer instead of stuffing more tokens into system prompts.

The project has been quietly compounding into something serious. The PyPI package describes it as an open-source memory backend for AI agents with REST API, MCP, OAuth, CLI, and dashboard, letting agents store decisions, share causal knowledge graphs, and retrieve context in 5ms without cloud lock-in or API costs. It ships under Apache 2.0, runs entirely on your own hardware, and speaks both the Model Context Protocol and a plain HTTP API.

Semantic memory, minus the vendor

The core is a semantic memory layer that gives assistants like Claude persistent storage across conversations and sessions. Content gets embedded with sentence transformers, and searches return semantically similar content rather than exact text matches, so a query for "authentication issues" also surfaces notes about "login problems" or "JWT errors."

What separates it from a bare vector DB is the surrounding machinery: a knowledge graph with typed edges (causes, fixes, contradicts), autonomous consolidation that compresses old memories as they decay, an