AI agent memory checks raised CLBench pass rate from 39% to 73%, a post reports

A post describing Microsoft research says memory curators risk preserving an agent’s mistakes, overgeneralizing from partial evidence and retaining stale facts when they rely only on its completed work. The approach it describes checks candidate memories against the live environment before saving them. In a GitHub Copilot harness on CLBench, the post reports pass rate rising from 39% to 73%, queries per question falling from 8.8 to 4.7 and task-agent cost dropping from $3.38 to $1.68. Across 90 consulting tasks in six environments, it says every memory configuration beat the baseline and tool calls fell by 16% to 75%.