Agent memory with no ingest LLM — and it beats mem0 on mem0's turf

What it does

nautilus-compass is an MCP plugin (Claude Code/Desktop, Cursor, Cline, Continue.dev, Zed, anything that speaks MCP) that bolts three things onto your agents. Long-term memory: raw text embedded locally with BGE-m3 — no extraction LLM at ingest, no graph, nothing leaves your machine. Drift detection: every prompt is scored against an anchor set of real failure patterns before the agent acts. And cross-agent contracts: when multiple agents or dialogs share files, it derives implicit contracts from handoffs, tracks closure, and audits for fake-closure. It runs as a local daemon; a hosted multi-tenant gateway is in open beta if you’d rather not self-host.

The interesting bit

The retrieval trick is almost embarrassingly simple: for single-session and knowledge-update questions, the answer usually lives in one user turn, and embedding the whole session dilutes it. So those question types route to turn-window chunks while everything else gets session-level hybrid BM25 + dense retrieval (RRF-fused). The README claims this flipped a −30-point deficit into a win over mem0 — and, refreshingly, publishes the experiments that failed (cross-encoder reranking hurts on this corpus; candidate-pool K is a no-op).

Key highlights

• LongMemEval-S full 500: P@1 0.890 / P@5 0.978 / MRR 0.929 vs mem0 2.0.19 at 0.774 / 0.916 / 0.834 — and it also overtakes mem0 on LOCOMO-10, mem0’s home benchmark.
• Drift detection: AUC 0.83 held-out, p95 latency under 50 ms, 0.5% fire rate in production traffic, anchored on 25 positive + 35 negative real failure patterns.
• Fully local and ~14× cheaper to reproduce: roughly $3.50 per 500 questions vs $50+ for GPT-4o-judged stacks.
• Cross-agent governance: contract derivation from handoff files, closure tracking, fake-closure/red-drift audits — with a 4-dialog, 28-hour field study in the repo.
• Ships 17 MCP tools with TLS and RBAC, plus an A2A adapter with mTLS and scoped peers.

Caveats

• Every head-to-head is the project’s own reproduction (each side on its own default embedder), not a neutral evaluation; there’s no mem0 comparison on LongMemEval-M yet, and the EverMemBench claim deliberately stops short of SOTA since OMEGA / Mem0g haven’t reported publicly.
• The LongMemEval-V2 numbers have a moving measurement story: an untuned first run scored 19.6% (web), a tuned re-judged run 40.0%, after a correction revealed the original judge was silently zeroing answers. Transparent, but still settling.
• Field-deploy notes admit rough edges: first recall after daemon idle can take up to 90 s on cold model load, a mis-set file-watcher flag used to silently hide fresh writes, and drift used to silently return “no risk” when its anchor file was missing (it now fails loudly).

Verdict

If you run several agents or dialogs over shared files and want memory that stays on your machine, this is one of the more evidence-dense options around — failed experiments included. If you want a white-box/graph memory or third-party-verified numbers, wait for the neutral benchmarks.