Graph memories fail on retrieval, not on storage: you stored something about a person under a name you no longer recall exactly, and exact-match lookup returns nothing while the data sits there. Semantic matching over node names solves the common case. Deleting a single observation rather than a whole entity is the other well-judged piece — a memory that was true and is now wrong should be corrected without discarding everything else known about that entity, and most implementations make you choose between keeping the error and losing the record.
A knowledge-graph memory server: entities hold accumulating observations, relations connect them, and retrieval is by fuzzy semantic search or by opening named nodes directly.
- search_nodes performs a fuzzy semantic search rather than exact matching, which is what makes a graph findable when you half-remember what you called something.
- Entities and relations are created in batches rather than one at a time, so building out a structure is one call.
- add_observations accumulates onto an existing entity rather than replacing it, which is how a memory grows without losing what was already there.
- Deletion is granular in three forms — whole entities with their relations, specific observations from an entity, or specific relations — so pruning does not mean starting again.
- read_graph returns the whole thing, and open_nodes fetches specific ones by name.
Nothing to supply.
One command — npx -y github:flrngel/fuzzy-memory-mcp#main