mem0-mcp
An archive for the things your AI should not forget.
01 · About
What it is.
A Dockerized MCP server that gives any Model Context Protocol client a long-term memory. It wraps mem0 and Qdrant into a single endpoint at http://127.0.0.1:8888/mcp — running entirely on your own machine, indexed per user and per project.
Every fact you store survives context-window resets, process restarts, and client switches. The database is a file in ./data/. Nothing leaves the host except calls to your configured embedder.
02 · Tools exposed
Four verbs.
add_memory(content, user_id, project)store a factsearch_memory(query, user_id, project, top_k)semantic retrievallist_memories(user_id, project, top_k)enumeratedelete_memory(memory_id)forget
03 · Install
Three lines.
git clone https://github.com/cocodedk/mem0-mcp.git
cd mem0-mcp && cp .env.example .env # add your OPENAI_API_KEY
docker compose up -dQdrant runs in the background. The MCP server listens on loopback 127.0.0.1:8888. That is not a mistake.
04 · Connect
From any MCP client.
{
"mcpServers": {
"memory": {
"url": "http://127.0.0.1:8888/mcp"
}
}
}05 · Context
Why bother.
Conversations end. Context windows close. But some facts — preferences, decisions, corrections, past incidents — deserve to outlast the session. This server keeps them in your own vector store, indexed by who you are and which project you are in, retrievable by any client that speaks MCP.
Security note. The endpoint has no authentication. It binds to 127.0.0.1 by design. Do not change that prefix. Do not expose port 8888 publicly. Do not run this on a shared machine without firewalling.