Commit c8df078
committed
feat(memory): HybridRetriever + progressive enhancement flags ship as v0.1.248
Cumulative Steps 2-7 rollout:
- Step 3 (GREEN-soft): HybridRetriever = BM25 + dense RRF + Cohere rerank blend,
shipping as the primary retrieval upgrade on top of dense-only baseline.
Phase B N=500 at gpt-4o measured 73.2% unweighted on LongMemEval-S.
- Step 2 (RED): SessionRetriever (rerank-skip root cause)
- Step 4 (RED): HyDE query rewriting (hypothesis replaces concrete tokens)
- Step 5 (RED): FactSupersession LLM post-filter (over-drops)
- Step 6 (RED): Split-on-Ambiguous rerank refinement (monotonic on wrong axis)
- Step 7 (RED): Observer/Reflector pipeline wired as additive overlay (dilutes retrieval pool)
All experimental features ship opt-in; HybridRetriever is the only production-grade addition.
Bench adapter (agentos-bench) measures every feature via matched-ablation + bootstrap CI + kill-criterion methodology. Umbrella findings and per-step docs in packages/agentos-bench/docs/.1 parent 3521e4a commit c8df078
0 file changed
0 commit comments