MCP Memory Service 🔌 MCP server Open source

What it does

MCP Memory Service gives coding agents a long-term memory that lasts between sessions. The agent can save a decision, a code pattern or an architecture note in one conversation and find it again later with semantic search. That saves you re-explaining the project every time. Memories are stored with tags and embeddings, which are computed locally with ONNX. They can be linked in a knowledge graph with typed relationships, and a consolidation job compresses older entries. The same service also exposes a REST API and a web dashboard. This lets several agents, or frameworks such as LangGraph and CrewAI, share one memory store. It is a Python project under Apache-2.0 with frequent releases.

Tools it gives your agent

• Core memory: memory_store, memory_search, memory_list, memory_update, memory_delete, memory_store_session
• Maintenance: memory_consolidate, memory_cleanup, memory_quality, memory_health, memory_stats
• Knowledge: memory_graph, memory_explore, memory_conflicts and memory_resolve for relationships and contradictions
• Ingestion: memory_ingest for documents and memory_harvest for pulling learnings out of sessions
• Mistake notes: add, search and update notes about past errors so the agent doesn't repeat them

Works with

By default it runs locally over stdio. An HTTP mode adds the REST API and dashboard, and a Streamable HTTP mode with OAuth supports remote clients such as claude.ai connectors. The README lists Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, Codex CLI, Gemini CLI, OpenCode (as a plugin), Goose, Zed, Amp and others.

How to connect

pip install mcp-memory-service
claude mcp add memory -- memory server

Cost and limits

It is free. By default it uses a local SQLite-vec store. Cloudflare, hybrid and Milvus backends are optional and can cost money; the changelog mentions one release that burned through a Cloudflare free-tier allowance. The default embedding model only handles English, and switching models means re-embedding what you already stored.

Safety notes

It binds to 127.0.0.1 by default. If you expose it with 0.0.0.0 or a tunnel, use HTTPS and OAuth. Don't use MCP_ALLOW_ANONYMOUS_ACCESS outside local testing. Recent releases fixed several serious auth advisories, so stay on the latest version. The agent can also delete memories, so back up the database.