Microsoft paper: read-only verification tools lift agent memory pass rate from 39% to 73%
dair_ai · x · 2026-09-11
Microsoft researchers tackle a core flaw in production agent memory: curators usually only read the finished trajectory, so they save mistakes, overgeneralize from partial evidence, and keep stale facts.
Method: give the curator a few read-only tools to verify each candidate memory against the live environment before saving. The task agent, retriever, and memory format stay unchanged — no retraining required.
Results:
- In a GitHub Copilot harness on CLBench, pass rate jumps from 39% to 73%
- Queries per question drop from 8.8 to 4.7; task-agent cost falls from $3.38 to $1.68
- Across 90 consulting tasks in six environments, every memory configuration beats baseline, with tool calls down 16-75%
Highly relevant for teams running persistent memory in production agents.
More from coding & agent
- ApprenticeBench: Agents Continually Learn Real Jobs, Surpassing Human Pros — ysu_nlp · 2026-09-12
- Agora open-sources meeting copilot demo powered by GPT-Live-1 — testingcatalog · 2026-09-12
- Open-source Agora meeting copilot puts GPT-Live-1 in your video calls — testingcatalog · 2026-09-12
- Ex-engineering manager: I now run Claude and Codex agents like I once ran dev teams — letandrewcook · 2026-09-12
- astra thrives on context: minimal prompting massively underperforms, dev finds — brandon_galang · 2026-09-12
- DeepSeek has no official coding harness; community Deep Code CLI hits 2.2k stars — PawelHuryn · 2026-09-11