Solo developer building nautilus-compass — a local-first memory layer for AI agents.
- Zero LLM calls at write time. Verbatim storage + local BGE-m3 embedding; all intelligence lives at read time (utterance-type routing, hybrid BM25/dense retrieval, date anchoring).
- Sealed benchmarks. Every number ships as a signed, byte-recomputable evidence pack — verify without trusting us.
- Head-to-head vs mem0 2.0.19 on LongMemEval-S (full 500, same questions & criteria): retrieval P@1 0.890 vs 0.774.
- Also: pre-action drift detection (AUC 0.83, p95 <50ms) and cross-agent contract tracking.
130 days, 771 commits — most of them landed by my own agent fleet. The tool is its own best demo.
Get it: pip install nautilus-compass | hosted open beta