Nautilus Compass

An open-source memory and reliability layer for AI agents: long-term memory that embeds text locally, drift detection that scores prompts before agents act, cross-agent contracts and governance, and an independent judging and verification pillar.

Raw text is embedded locally with BGE-m3 with no extraction LLM, no graph, and no data leaving your machine. Utterance-routed chunk retrieval handles single-session and knowledge-update questions, while session-level hybrid BM25 plus dense RRF covers the rest.

The README reports beating mem0 on all three LongMemEval-S metrics while staying fully local and 14x cheaper: P@1 0.890 vs 0.774, P@5 0.978 vs 0.916, and MRR 0.929 vs 0.834 on 500 questions. These are the project's own reproductions.

Before an agent acts, compass scores every prompt against an anchor set of 25 positive and 35 negative real failure patterns, with a reported held-out AUC of 0.83, p95 latency under 50 ms, and a 0.5% fire rate in production traffic.

It ships as a PyPI package and an MCP server with 17 tools, with plugins for Claude Code, Claude Desktop, Cline, Cursor, Continue.dev, and Zed. Retrieval reproduces with bash scripts/reproduce_lmes_retrieval.sh, or SUBSET=12 for a smoke run; CPU works and a GPU is recommended for the full 500.

A newer pillar: NACRE judge v1 compresses judging knowledge into a 1.7B LoRA with 88.5% three-state accuracy and ECE 0.072, caliber-bench scores benchmarks themselves, and a pre-registered deployment-precision study covers 291 cases, with negative results published as-is.

No. The digest's README covers memory, drift detection, judging, and benchmarks without mentioning Jev or TypeSafe; the repository entered this catalog through the cobanov community list.