v1.0.4 — Entity Graph, Trust Feedback, BM25 Hybrid

Added

• Entity graph: SQLite memory_entities + memory_feedback tables, deterministic regex entity extraction, auto-backfill
• 4 new tools: scope_recall_context, scope_recall_probe, scope_recall_related, scope_recall_feedback
• BM25 hybrid retrieval: FTS5 candidate pool now uses ORDER BY bm25() before recency tie-breaking
• Trust scoring: feedback rating (+1/-1) adjusts trust by +0.08/-0.12, applied in recall ranking
• Entity overlap bonus: memories sharing entities with the query get ranking boost
• Quality weight: trust/importance/confidence metadata factored into recall scores

Changed

• storage_views.py: FTS5 candidates BM25-ordered, configurable candidate_pool
• recall.py: applies quality_weight and entity_overlap_bonus from graph layer
• Extended stats with scoped entity and feedback counts

Fixed

• Block [System note] gateway wrappers and context-compaction handoffs from automatic capture

Full changelog