Move a document to a different location in the vault.

Read the full content of a document by id.

Search the memory vault. Returns top-k matching memorys (id, title, type, scope, summary, score). WHEN TO CALL (recall is part of acting, not a separate step): At session start (once): query for active-project + user-preferences to load durable context. Before writing/editing a file: query with a description of what you are about to write (e.g. 'creating React input with focus styles'). This catches lessons before mistakes. Before giving a multi-step plan or recommendation: query for preferences that shape format/scope. When the user's prompt touches a topic that may have a stored lesson, decision, preference, or project-fact. Before save_memory: query to avoid creating a duplicate. WHAT TO DO WITH HITS: READ `score_kind` FIRST — the bands below only exist on the fused scale. `score_kind: "rrf"` (fused rank sum, bounded): score >= ~100 with title/recall_when match: load_memory and apply the lesson before acting. score 30-100: read the summary, load if directly relevant. score < 30: usually noise; skip unless the summary is a perfect topic match. `score_kind: "bm25"` (also flagged `unfused: true`): the vector arm did not run — no embedding model, a cold-start timeout, or an open circuit breaker (`degraded` names which). These are raw MiniSearch scores on an OPEN scale — six figures on a real vault — so 100 means nothing here and every hit would look REQUIRED. Judge those hits by title, summary and recall_when match, and by their ORDER, never by the number. Never ignore a `lesson` hit with strong recall_when match. Two scores are comparable only within the same `score_arms` (and the same `score_version`): the fused scale reaches 163.934 with the two personal arms and 241.803 once the Bastra Commons contribute a third. A batch response whose phrasings disagreed on that reports `merged_by: "query-rank-fusion"` and drops back to `unfused` — its order is meaningful, its numbers are not a band. On the hybrid (BM25 + vector) path the score is a scaled rank sum, not a similarity — a top hit is high by construction. When the response carries top-level `weak_result: true`, no returned hit has a recall_when or title match: the high scores are likely rank-1-of-nothing, so prefer not to load them. recall returns lean CANDIDATES (no bodies). This is step 1 of a two-step flow: call load_memory ONLY for the hits you actually need — do not load every hit.

Change the category of a document.

Save a new document to the vault.

Persist a new memory into the vault as a markdown file with YAML frontmatter. This is YOUR long-term memory — save autonomously when a memory-worthy moment occurs, do not wait to be asked.

Save a product documentation entry.