Module recall 

Memory Recall Engine — Hybrid vector + FTS search (Phase N10).

Auto-embeds memories at write time using the Embedder trait, and fuses Tantivy BM25 + vector cosine similarity at recall time.

Architecture

RecallEngine
├── embedder: Arc<dyn Embedder>
├── vector_store: VectorStore
├── search_engine: SearchEngine (Tantivy BM25)
├── embedding_cache: HashMap<content_hash, Vec<f32>>
└── hybrid_search(query, limit) → Vec<RecallResult>
    1. Embed query → query vector
    2. Tantivy BM25 search → text scores
    3. Vector cosine similarity → vector scores
    4. Fuse: w1*BM25 + w2*vector + w3*importance
    5. Return ranked results

Environment Variables

Variable	Default	Description
WM_RECALL_BM25_WEIGHT	0.5	Weight for BM25 text score
WM_RECALL_VECTOR_WEIGHT	0.3	Weight for vector cosine similarity
WM_RECALL_IMPORTANCE_WEIGHT	0.2	Weight for memory importance
WM_TRUST_WEIGHT	0.0	Post-fusion trust multiplier (source_trust 0.7 neutral)
WM_RECALL_CONFORMAL_ALPHA	unset	Conformal set miscoverage level (unset = off)

Structs

BackfillReport

Report from a RecallEngine::backfill_embeddings pass.

RecallConfig

Configuration for the recall engine.

RecallEngine

Hybrid recall engine combining BM25 + vector search.

RecallResult

A single recall result with fused scores.