[FEATURE] RRF hybrid search fusion of FTS5 + ChromaDB

[github-actions[bot]]

Originally posted as issue #2000 by @thedotmack
#2000

Type: feature / enhancement request

Description: RRF hybrid search fusion of FTS5 + ChromaDB

---

Consolidated during the April 2026 backlog cleanup. Original: #1789

[devkancheti4-design]

Some data for this, measured offline with claude-mem's own search building blocks:

• FTS5, queried the way SessionSearch queries it;
• Chroma's default embedding, all-MiniLM-L6-v2.

The test set is LOCOMO's 1,536 natural-language questions (categories 1-4), with each dialogue turn a document and
one index per conversation.

retrieval	recall@10
FTS5, AND of all terms (today)	0.001
FTS5, AND first, then OR for 3+ terms (#4697)	0.534
Chroma default embedding (today's relevance path)	0.524
RRF (k=60) of Chroma + FTS5 (AND, then OR)	0.614

Fusion adds +0.090 recall@10 (about 17%) over Chroma alone on this set. With today's AND query, fusion adds
nothing (0.525), because FTS5 returns no candidates for questions. #4697 fixes that part first.

A small, opt-in proposal:

• a pure rrfFuse(lists, k = 60) helper;
• a setting such as CLAUDE_MEM_SEARCH_FUSION=rrf, default off;
• on the relevance path only: query Chroma and FTS5 50 deep each, fuse, then hydrate as today. Date-ordered and
  filter-only searches stay unchanged.

I'm happy to send it as a narrow PR with tests and this eval, if that's a direction you'd take. Caveat: the documents
are raw dialogue turns, not compressed observations.

The script
reproduces every number above.