MaxGraph

What MaxGraph MCP server does

The MaxGraph MCP server adds persistent memory to AI-agent workflows. It records project facts such as decisions, bugs, tasks, and agreements, including their dates, so later sessions can retrieve relevant context instead of starting from an empty conversation. Stored records remain separate from chat transcripts, so deleting or closing a conversation does not remove the project memory.

The system combines a knowledge graph with hybrid search. It can provide a short project summary at the beginning of a session and surface unfinished work without requiring the user to restate it. A local dashboard shows the stored information through a project memory map, search interface, and event timeline.

Use this when continuity across sessions matters, particularly for ongoing software projects or agents that need to recall prior decisions. The material describes support for Claude, ChatGPT, Cursor, Kimi, DeepSeek, and other agents that can connect external tools.

How it works

MaxGraph runs as a self-hosted Docker deployment. Agents connect to the streamable-HTTP MCP endpoint, normally at http://host:8787/mcp, using a Bearer token. ChatGPT and OpenAI-compatible systems can use the REST API or an agent adapter, while other HTTP-capable agents can connect through REST.

The MaxGraph MCP server keeps records on the user’s computer as ordinary text files under the data/ directory. These files can be opened, read, copied, and backed up independently of the chat client. The dashboard is available at /app; a local installation therefore uses http://localhost:8787/app.

Setup and configuration

Docker is the only stated prerequisite. Clone the repository and start the included Compose deployment:

git clone https://github.com/RuMemory/maxgraph
cd maxgraph
docker compose up -d --build

The first start generates an access token. The token can be read from the container logs or from data/local_token.txt, then entered in the dashboard and supplied to the client’s MCP configuration. Provider-specific examples are available in the project documentation.

The deployment checks for updates once per day by default. Set AUTO_UPDATE=0 to disable that behavior. No required environment variables are documented for the initial Docker setup.

Tools and capabilities

The MaxGraph MCP server provides the following documented capabilities:

• Persist project facts across agent sessions.
• Retrieve information through hybrid search.
• Represent project memory as a knowledge graph.
• Generate a concise project context summary for new sessions.
• Surface unfinished work and prior agreements.
• Expose stored memory through an MCP streamable-HTTP endpoint and REST API.
• Display memory, search results, and events in a web dashboard.

Limitations and notes

The repository uses the Business Source License 1.1. Personal and internal use is free according to the README; selling MaxGraph as a service requires a commercial license. Each version becomes MIT-licensed four years after publication.

The README describes a seven-day trial. After the trial, only writing new facts stops; the dashboard, search, and export remain available. The service is self-hosted, and the documented storage model keeps records on the local computer rather than in the chat transcript.