marm-memory MCP
A local-first MCP memory server for agents: session history, codebase index and a concept graph in one SQLite database
- 438 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- Indexes your project code and history on your machine
- Runs a local server and stores a SQLite database
Install
- Works in
Manual install
pip install marm-mcp-serverLocal stdio run, needs Python 3.10.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Indexes your project code and history on your machine
- Runs a local server and stores a SQLite database
This is third-party code. Review the repository files before installing.
What it does
The server gives agents long-term memory that lives on your machine without the cloud. It fuses three things: session history, a codebase index and a graph of links between concepts, and keeps it all in SQLite. An agent reaches the memory through plain phrasing rather than direct tool calls and gets fast access to the context it needs. It works with Claude Code, Codex, Gemini, VS Code, Cursor and other clients. It runs locally from PyPI over stdio or as a Docker container with HTTP access, and supports several agents at once.
Who it is for. For developers who want private agent memory with a code index and history, without the cloud.
Good fit when
- You want the agent to remember past sessions and decisions
- You want a local codebase index for fast context lookup
- You cannot send context to the cloud for privacy reasons
Not a fit when
- You need shared cloud memory for a distributed team
- You are not ready to run a local database and server
Example request
Recall what we decided about the database schema in the last session of this projectMCP
- Transport
- stdio, http
- Authentication
- not required
Environment variables
MARM_API_KEYsecret- Key to protect access in Docker HTTP mode
Limitations
Memory is stored locally in SQLite and does not sync across machines by itself. It needs Python 3.10 and, for the HTTP mode, a Docker run where access can be protected with a key. The server indexes your code and history on your machine, so it does more than read instructions.
How to disable. Remove the marm-memory server from your client's MCP config and uninstall the package: pip uninstall marm-mcp-server or stop the Docker container. The SQLite database can be deleted manually.
FAQ
Does data go to the cloud?
No, memory lives locally in SQLite on your machine; the project is built for privacy.
How do I connect over HTTP?
Run the Docker container and add the server to your client at a localhost URL, optionally protecting access with MARM_API_KEY.
Discussion
No questions or reviews yet.
Related
MCP serversai-memory keeps long-term assistant memory in local SQLite and serves it over MCP, an HTTP API and a CLI
LongMemory keeps project memory in local SQLite and serves it to an agent over MCP: decisions, task state, skills and a code graph
A self-hosted MCP server for long-term agent memory: structured records, versioned rollback and a visual dashboard
Long-term memory for agents: a remote MCP server, plugins for Claude Code, Cursor and Codex, and an API for your own apps
iai-memory
iai-memory (Personal Memory Engine)
A local long-term memory MCP server for agents: keeps conversations verbatim and returns the right context at the start of a session
An MCP server for long-term agent memory: compresses dialogues into atomic facts, resolves pronouns and anchors relative time to absolute dates
Comparison with similar tools
Cards that share the main search query, and how they differ.
README in short
The README presents a local memory server for agents that fuses session history, code indexing and concept graphs in SQLite and works without the cloud. Installation is via pip install marm-mcp-server or Docker, with stdio and HTTP transports, ready config snippets for clients and a claude mcp add command. It documents a suite of sixteen tools, scaling benchmarks and an approach where you talk to memory in plain words rather than calling tools. The project is published in the MCP registry, on PyPI and Docker Hub. Apache 2.0 licensed.