Compass v0.9 · LongMemEval-S 56.6% · cross-agent memory federation 2026-05-05 · for HN / 知乎 / X / weibo · 1500 字 · 草稿 TL;DR We achieved 56.6% on LongMemEval-S (n=500) with DeepSeek V3.2 + local bge-m3 + a 5-component pipeline · matching the Zep SOTA band at 1/15 the cost. The plugin (Compass v0.9) ships an MCP server, A2A adapter, npm wrapper, and one-line Nautilus agent integration. The killer feature isn't the accuracy. It's cross-agent memory federation : same user_id across Claude Desktop, C
Comment
Sign in to join the discussion.
Loading comments…