Cleaning an Agent's History Hurts When the World Still Holds the Error
Source
arxiv.org
Author
Egor Pakhomov, Anurag Koul, Hongjun Liu, Bo Pang, Shafiq Joty, Erik Nijkamp
Date
Why it matters
If you compact or hand off agent histories, how the correction is phrased matters: a note beside the preserved error helps, while rewriting it as a completed action hurts recovery when the world still holds the error.
Key takeaways · AI-distilled
The authors freeze a tool-using AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → mid-task with its exact world state, then edit either its history or the world and let a separate receiver model continue, so a history edit can be tested against what the world actually holds.
On AppWorld, a database write's receipt reporting success lowered whole-task success from 45% to 28% when the write's effect was absent, but raised it from 38% to 49% once the effect was present.
Pooled over two receivers, those two receipt effects differ by 27 points; the authors report the difference is positive in six of eight receiver models at 97.5% and in seven at 95%.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
articleThe Memory Trust Gap: Capability-Dependent Failures in Persistent-Memory AgentsJundong Hu, Shekar Ramachandran
articleKnowing When to Stop: The Evidence-Action Gap in Coding AgentsTianyu Chen, Xinran Zhao, Vijay Viswanathan, Shoubin Yu, Mingyuan Zhou, Tongshuang Wu
articleCompact-Bench: Measuring What Context Compaction Costs Long-Horizon Coding AgentsRuopei Sun, Jianfeng Cai, Kangwen Zhao, Wengang Zhou, Houqiang Li