claude-flow

First implementation landed from the dream-cycle research cluster (#2316-#2324). Adds entity matching as a third RRF arm in hybridSearch alongside dense (HNSW/RaBitQ) and sparse (FTS5/BM25), plus per-result signal provenance.

What's new

@claude-flow/memory 3.0.0-alpha.20 — entity arm + signal provenance in the hybridSearch controller:

• entity-tagger.ts — regex extractor for emails, URLs, file paths (POSIX + Windows), quoted phrases, proper-noun 2-grams. Deliberately conservative: false negatives OK, false positives would dilute RRF.
• hybridSearch now runs three arms in parallel: dense + sparse + entity (per-token keyword scan, gated on extractEntities(query).length > 0). Empty entity set drops the arm rather than passing [] to dilute fusion.
• signals: ('vector' | 'bm25' | 'entity')[] on every fused result. Computed by pre-fusion set membership; lets callers debug which arms surfaced an entry without re-running the search.

Capability smoke (end-to-end against built dist)

Corpus: 30 generic "authentication" entries + 1 "Alice Smith" needle. Query: "Alice Smith authentication":

score=0.0477  signals=["vector","bm25","entity"]  key=alice-needle      ← #1
score=0.0323  signals=["vector","bm25"]           key=generic-1
score=0.0323  signals=["vector","bm25"]           key=generic-0
score=0.0313  signals=["vector","bm25"]           key=generic-3
score=0.0301  signals=["vector","bm25"]           key=generic-2

Alice ranks #1 with full triplet provenance — runners-up only fire on vector + sparse. ~47% RRF score boost from the entity signal.

Packages

| Package | Old | New | Tags |
|---|---|---|---|
| @claude-flow/memory | 3.0.0-alpha.19 | 3.0.0-alpha.20 | latest, alpha, v3alpha |
| @claude-flow/cli | 3.10.38 | 3.10.39 | latest, alpha, v3alpha |
| claude-flow | 3.10.38 | 3.10.39 | latest, alpha, v3alpha |
| ruflo | 3.10.38 | 3.10.39 | latest, alpha, v3alpha |

@claude-flow/cli's @claude-flow/memory dep pinned to ^3.0.0-alpha.20 so wrapper users get the entity arm automatically. v3/pnpm-lock.yaml regen included (lesson from #2311 — bumping a workspace dep without lockfile regen breaks pnpm install --frozen-lockfile).

What this implements vs the dream-cycle ADR

ADR-147 (#2317) split the work as P1 "wire FTS5 + RRF fusion" and P2 "entity arm + provenance". The investigation found P1 was already shipped in controller-registry.ts:713 before the ADR was filed — applyRRF(k=60) + applyMMR(λ=0.7) over dense + sparse was already in. This release lands the actual gap, P2.

Tracking note for the dream-cycle process posted on #2324.

Tests

• 12 new entity-tagger.test.ts (regex pinning — generic prose returns empty, and/or → empty, "a" over "b" → empty, single capitalized words → empty)
• 2 new graceful-retrieval.test.ts ADR-147 assertions (signal provenance on every fused result; needle-in-haystack)
• Full memory suite: 416/420 (4 pre-existing Windows-env failures in agent-memory-scope, auto-memory-bridge, benchmark — untouched files)

Out of scope (follow-ups)

• Dedicated SQL entity index — current per-entity searchKeyword calls are fine for typical query entity counts (1-3); unbounded if a query mentions 20+. A future ADR can add an entity_index table for hard-bound latency.
• Async writes by default (ADR-147 P3) — orthogonal; consolidator already handles HNSW background rebuild.
• LoCoMo benchmark publication (ADR-147 P4) — needs harness wiring + dataset access; separate workstream.