Nautilus Compass
Give agents drift-aware, cross-session memory over MCP/A2A so long-running solo projects retain context without blowing embedding budgets.
Overview
Nautilus Compass is a Build-phase MCP server that provides drift-aware cross-agent memory over MCP/A2A for long-horizon agent projects.
What is this MCP server?
- Drift-aware cross-agent memory exposed via MCP and A2A (DMAS-oriented)
- Publisher benchmarks: LongMemEval-S 56.6% and ~1/15 cost vs Zep (per registry metadata)
- Python PyPI package nautilus-compass v1.6.2 with python3 -m nautilus_compass.mcp_server
- Tags include BGE-M3, longmemeval, and agent-infrastructure in registry metadata
- Research paper linked at arxiv.org/abs/2512.02410 for methodology depth
- Version 1.6.2 on PyPI
- LongMemEval-S 56.6% (publisher benchmark)
- Cost vs Zep stated as 1/15 in registry metadata
Community signal: 4 GitHub stars.
What problem does it solve?
Agents forget prior decisions across sessions, and cheap memory layers drift without you noticing until outputs contradict your shipped product.
Who is it for?
Solo builders running several MCP agents on one product who need durable, drift-checked recall without enterprise memory pricing.
Skip if: One-shot codegen tasks, teams that forbid external memory services, or workflows needing only ephemeral chat context.
What do I get? / Deliverables
Multiple agents share a monitored memory layer with published LongMemEval-S performance and lower stated cost than Zep-style stacks.
- Cross-agent retrievable memory graph
- Drift-aware recall for multi-session agent work
Recommended MCP Servers
Journey fit
Build is the canonical shelf because memory infrastructure is adopted when assembling agent tooling, even though recall pays off through operate and grow. Agent-tooling is where MCP memory servers plug into Claude Code, Cursor, and multi-agent setups as shared recall layer.
How it compares
Specialized agent memory MCP with drift awareness, not a general-purpose SQL database or note-taking plugin.
Common Questions / FAQ
Who is Nautilus Compass for?
Indie developers and small teams orchestrating multiple AI agents who need shared, long-horizon memory via MCP rather than copy-pasting context each session.
When should I use Nautilus Compass?
Use it when project memory spans days or weeks—specs, customers, and architecture—and you want drift detection plus MCP/A2A integration from the build phase onward.
How do I add Nautilus Compass to my agent?
Install nautilus-compass from PyPI, configure stdio with python3 -m nautilus_compass.mcp_server in your MCP host, then follow Compass docs at compass.nautilus.social for memory write/read policies.