About Nautilus Compass
Compass is an open source memory and reliability layer for AI coding agents. It stores long term memory locally, embedding raw text with a local BGE m3 model rather than calling an LLM to extract facts, so data stays on the user's machine. It connects through MCP and A2A and lists support for Claude Code, Claude Desktop, Cline, Cursor, Continue.dev and Zed, with a quickstart script that generates a token and writes the MCP configuration.
Beyond memory, Compass offers drift detection, which scores an agent's planned action against known failure patterns before it runs, and cross agent contracts that infer implicit agreements between agents sharing files and track whether they are closed. The vendor publishes benchmark results on LongMemEval and LOCOMO against mem0, along with its failed experiments, and says the work is reproducible. The site is written in Chinese.
A second part of the platform is an independent evaluation service covering a harness leaderboard, a grading model and a meta benchmark that tests evaluations themselves.
Pricing as published is limited to that service. Submitting a harness for a basic review is free and returns a report card within 24 hours, a deeper report costs $199 with a 48 hour turnaround, and a custom review is also offered. The software is released under a modified MIT license.