Smarter and faster search, measured in public.
- Natural-language questions no longer dead-end: when the exact ladder finds nothing, a relevance tier ranks sessions by informative-term overlap — with tf saturation, a same-message co-occurrence boost, and single-term candidates strictly behind multi-term ones. Quoted phrases keep exact semantics; single-word noise stays gated.
- Published benchmarks, harnesses in-repo: LongMemEval-S session retrieval 84.9% R@1 / 94.0% R@5 (cleaned 470), LoCoMo 63.7% R@1 / 79.5% R@5. No LLM, no embeddings; run them yourself in minutes.
- Performance: session-start hook ~50ms warm (60s digest cache per cwd), MCP recall ~14ms warm (manifest cache), cold index 21s -> 12s (parallel parsing and redaction).
- Deterministic ranking on ties.