mem-agent-mcp: local memory for an agent

Install

Works in

• Claude Code
• Claude Desktop
• SourceCraft
• Koda
• Coddy

Manual install

make install
make setup
make run-agent
make generate-mcp-json

What it does

The server exposes one tool, use_memory_agent: the agent asks a question, and Dria's local mem-agent model searches a folder of notes and returns the answer. The notes are Markdown files linked in the Obsidian manner. The main file, `user.md`, describes you and your relationships, and every entity, whether a person or a company, gets its own file in the `entities` folder. You can edit the notes by hand without restarting the server. The model runs on your computer. On macOS it runs through LM Studio with the Metal backend, and on Linux with a GPU through vLLM. On ARM64 Linux vLLM is not installed by default; there the model is reached through OpenRouter, OpenAI or a LiteLLM proxy. The model comes in 4-bit, 8-bit and bf16 variants. You can fill the memory with ready connectors. They parse a ChatGPT history export, Notion and Nuclino exports, and read GitHub repositories and Google Docs folders through their APIs. The `make memory-wizard` helper guides the setup. Answers can carry filters, for example a ban on revealing exact age or email addresses; filters live in a `.filters` file and are picked up on the fly.

Who it is for. For people who keep a personal Markdown notes base and want to query it from Claude or another client without handing the notes to an outside model.

Example request

Ask the memory what I noted about Acme Corp and which of my contacts handles purchasing there.

MCP

Transport

stdio, http

Authentication

not required

Limitations

macOS with the Metal backend and Linux with a GPU are supported. The mcp_server folder's `pyproject.toml` specifies Python 3.11. Before starting the server you must install LM Studio or vLLM, pick the memory folder with `make setup` and start the model with `make run-agent`: without a running model the server cannot answer. On ARM64 Linux an external model or a proxy is required.

The memory folder must follow the structure with a `user.md` file and an `entities` directory, and links between files must use the strict double-bracket format. The GitHub and Google Docs connectors need an access token, and a Google token lives for about an hour. The repository has not been updated since November 2025.

Whether Hugging Face and ChatGPT services are reachable from Russia was not checked for this card.

How to disable. Remove the mem-agent entry from the client configuration, in Claude Code that is `claude mcp remove mem-agent`, and stop the model. The notes folder stays on disk and belongs to you. If you exposed the server through a tunnel, close that too.

README in short

The document briefly introduces the server, lists supported platforms and walks through installation: `make check-uv`, `make install`, `make setup`, `make run-agent` and `make generate-mcp-json`. It then describes the memory structure with examples of `user.md` and entity files, filters, and a connector table with commands and steps for getting exports and tokens. Separate sections explain connecting Claude Code and ChatGPT, and the end gives troubleshooting tips and a Python skeleton for a new connector.