Summary
Memory you can inspect, not a vector blob.
Facts go in as named entities, typed relations and one-fact observations, which means you can read the graph back and see exactly what the assistant believes — and delete a single wrong observation without disturbing the rest.
What it is
An MCP server implementing persistent memory as a local knowledge graph on the Kuzu embedded graph database. Memory is modelled as entities with a type, relations written in active voice, and observations that hold one fact each.
What you get
- Entities created with a name, a type and a list of observations
- Relations created between entities in active voice, with duplicates skipped
- Observations added to and removed from existing entities
- Entities and relations deleted
- The whole graph read back
- Nodes searched by query, and named nodes opened in full
- Tags for cross-cutting classification of entities and observations, with an optional category and description
Requirements
The graph lives in a local Kuzu embedded database, so it stays on your machine. Entity names are the identifier: creating one that already exists is ignored rather than duplicated.
Setup effort
One command — npx -y @deanacus/knowledge-graph-mcp