[DottytheHomeless]

What

1. Conversation-search blind spot: search_misses showed 42 misses/7d, all source='search_conversations' with hit_count=0 — the conversations table has 0 rows (nothing ingests it), so daily-review flows were flying blind. searchConversations() now falls back to memory recall (marked source='memories_fallback', note='conversation_search_fallback: memories'); a miss is recorded only when the fallback also returns nothing.
2. Cross-layer dedup: the markdown importer (migrate-claude-memories.mjs) deduped by summary only for source='manual' rows, so native store_memory rows were invisible to it → markdown/native duplicate drift. Dedup now matches active memories by content_hash/exact content across all sources; summary matching kept only for legacy importer idempotency.

Evidence

• Pre-fix read-only check on live DB: [{"source":"search_conversations","count":42,"zero_hits":42}], conversations=0
• Write-time near-dup in index.mjs was already provenance-independent; the gap was importer-side only

Tests

node conversation-fallback-dedup.test.mjs — 5 passed / 0 failed (fallback hit, fallback metadata, no-miss-on-fallback-hit, importer idempotency, provenance-free skip)

🤖 Generated with Claude Code