Compass v0.9 · LongMemEval-S 56.6% · cross-agent memory federation

Short summary

Compass v0.9 hits 56.6% accuracy on LongMemEval-S by combining multi-angle query rewriting with per-model thinking optimization. The standout feature: cross-agent memory federation—same user_id works across Claude Desktop, Cursor, and Cline, so knowledge learned in one client instantly transfers to another. Install via pip/npm; deploy for $3.50/run; roadmap includes E2EE, RAID-2 review, and v1.0 in early 2027.

• •56.6% accuracy on LongMemEval-S—1/15 the cost of paper SOTA tier
• •Cross-client memory federation: learn once, recall everywhere across Claude Desktop, Cursor, Cline
• •Open-source with reproducible benchmarks; production deployment from $50-500/month per user

Generated with AI, which can make mistakes.