The rule the docs push hardest is to follow every store with a relationship - solution solves problem, fix addresses error. That is what turns the store from a pile of notes into something that answers why. The core profile is the default deliberately: the removed tools were dropped to keep the tool list from eating the context window.
A graph-backed memory store. Flat notes keep entries isolated; here memories carry typed relationships, so asking what happened with retry logic returns the causal chain rather than three unrelated notes.
- Recall as the first move - fuzzy matching with relationships included, tuned for natural-language questions
- A stricter search for exact matching, tag filters, importance thresholds and pagination when recall is too loose
- Storing typed memories - solution, problem, error, fix, decision, pattern - and reading one back in full
- Relationships between memories, with over 35 types across causation, solution, context, dependency, knowledge, comparison and workflow
- Exploring what a memory connects to, updating it, and deleting it
- Recent activity and store statistics for an overview
Bun, then install from the ts directory and link the binary; data lives in ~/.memorygraph. The MCP surface ships in two profiles: a core profile of 9 tools that covers memory CRUD, relationships and discovery, and an extended profile of 12 that adds database statistics, complex queries and contextual search. Projects share one store by default - set a different path per project, or tag memories with the project name and filter on it.
One command — pipx install memorygraphMCP