The isolation is the reason to run this yourself rather than use a hosted memory service — user, agent and session are separate namespaces, not tags on a shared pool. The provider choices are all configuration, so the same server runs on Redis and OpenAI or on local Ollama models.
A self-hosted MCP server that puts Mem0's AsyncMemory API behind MCP tools. It stores what an agent should remember, retrieves it by semantic similarity, and keeps memories scoped per user, agent and session rather than pooled.
- add_memory — store information in long-term memory with semantic indexing
- search_memories — retrieve by semantic similarity, with filters and a result limit
- get_memory — one memory by ID
- update_memory — change stored content
- delete_memory — remove it
- list_memories — list with filtering and pagination
- Multi-tenant isolation across user, agent and session scopes
- Two transports: stdio for a local agent, SSE for remote connections
Python with `uv`, or the included Docker Compose setup. Configuration comes from `~/.config/mem0-mcp-server/settings.json`, validated with Pydantic, choosing the vector store, LLM and embedder providers. `OPENAI_API_KEY` is required for the LLM in the documented default. `MCP_TRANSPORT` selects `stdio` or `sse` and defaults to `sse`, with `MCP_HOST` and `MCP_PORT` for the bind address. Settings resolve in order: tool parameters, then `MCP_`-prefixed environment variables, then the config file, then defaults. A Helm chart is included for Kubernetes.
One command plus a key — uvx mem0-mcp-server, then supply credentials