Memory recall
Find relevant active memories across the permitted owner scope, rather than searching only recent writes.
Candidate selection and ranking
Recall combines vector candidates with a bounded literal-match branch. It uses stored bge-m3 vectors where available and bounded fallbacks when vectors or SQL search are unavailable. Owner, scope, status and validity filters apply before ranking.
Ranking combines similarity, bounded intent boosts and outcome evidence. Age and confidence do not enter this ranking. Confirmations and distinct corroborating sessions can improve rank; refutations can lower it. Being served or unused does not improve rank.
Recall before a task
Pass the task about to be done in task (Companion: recall --for-task, and every phase brief). The recall then also reads what other agents of the project recorded under their own name, examines 60 candidates instead of 20, and keeps a memory when it applies to the task rather than when it answers a question. A recall without task is unchanged. The agent scope is a namespace chosen by the caller, not an access boundary.
Journal and history
Only active memories are recalled by default. Journal context and todos are excluded unless requested through include_journal or the journal type. Use include_inactive or memory_view: "history" to inspect retired entries and their retirement records.
Optional rerank and judge
Memory reranking and judging have separate switches from document retrieval. When a judge answers, it can return no_supporting_memory: true with an empty list. Provider failure retains the prior order. Read the data-location policy.
Response contract
Compact results retain memory IDs, content, source pointers, scores, evidence and retirement fields. With a judge, the default cap is three memories. Set include_metadata: true when full owner, authority or temporal metadata is needed.
snipara_recall({ query: "Why did we choose this index?", limit: 5 })Last updated: September 24, 2026