Search & retrieval

Three retrieval modes, one philosophy: show your work.

recall — the everyday search

recall("how do we handle rate limits?", project="myapp", limit=5)

Semantic search over memories and decisions (or keyword matching without the [semantic] extra — same API, honest downgrade). Results carry:

• score — similarity, already adjusted for decay (Beliefs, staleness & decay);
• belief — annotation if the memory is stale/superseded/low-confidence;
• scope — project vs global evidence (Projects & scoping);
• decay — if ranking was reduced by staleness, you see by how much.

recall(neighbors=true) — graph-aware search

Each result also brings its strongest typed graph neighbours (1 hop): what's connected surfaces even when it isn't textually similar. The neighbour list shows the relation (cites, supersedes, similar…), so you know why it appeared.

recall_associative — spreading activation

The deep cut. Your query's best matches become seeds; activation spreads along citation, semantic and entity edges (personalized PageRank). Memories strongly connected to the topic emerge even with zero textual overlap. The response includes the plain-cosine baseline so the difference is auditable — you can see exactly what the graph added.

expand_memory — from pointer to content

Dense results are pointers by design. expand_memory(ids=[12, 31, 44]) returns full contents in one call (max 10), and counts as an access — rejuvenating those memories against decay.

Semantic vs keyword

brain_status tells you which mode is active. Everything else — beliefs, graph citations, scoping, decay — works identically in both.