MCP Tools

Once the engine is wired into your editor (see Installer & doctor), these tools appear in the client’s tool list. Tool names use underscores (memory_observe, not memory.observe).

The one you’ll use most is memory_observe — feed it raw text and the engine decides what’s worth keeping. A good default is to call it after every user message.

Core

Tool	Purpose
`memory_observe`	Feed raw text; the engine extracts and saves what’s worth keeping. The primary tool — call it after every user message.
`memory_save`	Write an explicit memory item (scope, type, content, importance).
`memory_search`	Search stored memories by scope/project + content match.
`memory_recall`	Fetch a single memory item by id (reinforces it on access). Not a semantic search.
`memory_list`	Enumerate the most-recent items in a scope.

Lifecycle & correction

Tool	Purpose
`memory_extract_pending`	Return candidate memories awaiting confirmation.
`memory_commit_extraction`	Confirm and persist extracted candidates.
`memory_supersede`	Replace a stale memory with a corrected version (keeps provenance).
`memory_delete`	Delete a memory item.

Prediction & context

Tool	Purpose
`memory_predict`	Forecast the next likely event for a subject, with calibrated confidence.
`memory_build_context`	Assemble a context bundle (memories + patterns + predictions) for a subject.
`memory_prefetch_related`	Spreading-activation retrieval — surface memories linked to the current working set.

Knowledge graph

Tool	Purpose
`memory_query_graph`	Query entities and edges in the knowledge graph.
`memory_graph_reason`	Multi-hop reasoning over the knowledge graph.

Diagnostics

Tool	Purpose
`memory_stats`	Counts and aggregates for the store.

The exact tool set depends on your engine version. Always trust the tool list your connected client reports.