Memory MCP Server

Memory MCP Server is one of the reference server implementations maintained in the Model Context Protocol servers repository. It gives an assistant persistent memory across conversations by storing what it learns as a small knowledge graph rather than as a growing blob of text. The graph has three primitives: entities, which are named nodes with a type such as person, project, or organization; relations, which connect two entities in one direction and are written as verbs; and observations, which are short discrete facts attached to a single entity. Because facts live on entities as separate observations, memory can be edited surgically. An assistant can add an observation when it learns something, delete one that has gone stale, or remove a relation without disturbing the rest of the graph. Reading is equally granular: the server can return the whole graph, search nodes by a query that matches names, types, and observation text, or open a specific set of nodes by name when the assistant already knows what it is looking for. The graph persists to a file on the machine running the server, so it survives restarts and stays local. It is most often used as the memory backend behind a personal assistant setup, paired with a system prompt that tells the model when to recall from the graph and when to write to it. As a reference implementation it is deliberately small, which makes it a good thing to read, run locally, and fork into something backed by a real database once the pattern proves useful.