Memory

One of the active reference servers in modelcontextprotocol/servers. Stores a small knowledge graph to a single JSON-L file by default, so cross-session memory survives agent restarts without a database. Fits workflows that need structured long-term state — user profiles, project facts — more than fuzzy semantic recall. What it does: - Creates entities with a stable name and a free-form type ("person", "project", "decision") - Creates directional relations between two entities with a typed edge label - Adds free-text observations to any entity and queries the graph by name or relation - Retrieves the neighbourhood of an entity so the agent can walk connections rather than scan - Soft-deletes entities and relations so a full session can be audited after the fact - Pluggable storage: in-memory for tests, JSON-L on disk by default, custom driver via a small interface Example: Launch alongside a coding agent. As the agent works it adds "authored_by" edges between commits and team members and "touches" edges between commits and files. A week later a new question like "who last edited the auth middleware" resolves as a single neighbourhood query. Setup: No required env vars. MCP_MEMORY_FILE_PATH overrides the default storage location when you want to share the graph across two clients. Transport: stdio only. Notes: No embedding model, so retrieval is exact-match rather than semantic. Pair with an embedding MCP if you need similarity search over free-form notes. The on-disk format is plain JSON-L, so a human can grep or diff it — which is useful for debugging and dangerous for sensitive data. Treat the file as you would a small database: back it up, do not commit it.