Give your agent a memory that lasts with the Memory MCP server
The official Memory MCP server keeps a small knowledge graph in a local file. How to add it to Cursor and Claude, where to store the file, and what I saw testing version 2.1.1.
Most agents forget everything when the chat ends. The Memory server is the reference fix from the Model Context Protocol project. It stores what the agent learns as a small knowledge graph in a local file, and gives the agent tools to add to it, search it and clean it up. Nothing leaves your machine, and there is no account or key.
It shipped as version 2.1.0 in the v2.3.0 release of the reference servers on November 4, and npm now has 2.1.1. I ran 2.1.1 locally before writing this, and the notes below come from that run and the official README.
How the memory is shaped
There are three pieces. Entities are the things worth remembering, each with a name and a type such as person, project or company. Relations link two entities in active voice, for example Maya works_at Acme. Observations are single facts attached to an entity, kept one fact per line.
That structure is the point. Instead of dumping chat history into a file, the agent writes short facts it can find again later by name or by search.
Add it to Cursor or Claude
The server runs locally through npx. It needs Node.js 24 or later. Put this in your client's MCP config, with MEMORY_FILE_PATH pointing at a file you own:
