route-by-query-shape

Route relationship-shaped queries to a knowledge graph backend instead of vector search. The route-by-query-shape interceptor detects questions like "who did I talk to about X" and redirects memory_search calls to the graph where entity relationships are stored.

Vector search matches chunk text by semantic similarity, but answers to relationship questions live in entity relationships rather than text chunks. A knowledge graph backend stores these connections explicitly and resolves them correctly.

No. It only fires when the query matches a relationship shape, and it is capped at 12 activations per session. Non-relationship queries continue to use the standard vector-based memory_search backend.

The built-in interceptor lives in src/agents/skills/builtin-interceptors/route-by-query-shape.ts. It watches memory_search calls, extracts the salient entity, and modifies the call to target the knowledge_graph backend before execution.

Graph routing only helps for relationship-shaped queries about entities and their connections. For general semantic recall of notes, documents, or conversations, the standard vector-based memory_search remains the appropriate backend.