In plain words
Coding agents can look up records of earlier work, but a record may be too long or arrive at the wrong time to help. This paper introduces REM, a way to organize each past coding run into a short guide and show its detailed edits only after the current agent’s first failed check.
On 99 bug-fixing tasks, REM resolved a larger share than giving the agent no memory across each of three runs. The difference was not statistically significant in the paired test, so the result is promising evidence from this test set, not a guarantee.
The research
When an agent retrieves a relevant past solution, how can it avoid being distracted by too much detail or by advice that arrives at the wrong point in its work? The agent’s needs change during a task, while a retrieval system alone cannot observe that progress.
REM organizes past coding runs into short guides and reveals detailed edits only after a failed check, managing memory as the agent works.
The authors compared REM with no memory on 99 repository bug-fixing tasks, running each condition three times with Claude Sonnet 4.5. A task counted as resolved when the benchmark’s tests passed on the submitted patch.
REM resolved 27.6% on average, versus 24.9% without memory, and resolved more tasks in every replicate.
REM also used fewer cost-weighted tokens per task. But the paired resolution difference was not statistically significant. The authors therefore treat the consistent direction across runs as evidence that the lifecycle may help in this setting, rather than as a conclusive resolution gain.
The main resolution contrast was not statistically significant in a paired test, and the five-model sweep used one run per model. The results cover one repository bug-fixing test set and a fixed pool of prior runs, not memory that accumulates across repeated runs.
Analysis
Deep
- Original paper authors
- Xinyi Lisa Chen, Pedro Herrero-Vidal, Xiaohan Chen, Xiaojing Yu, Cris Liu
- Submitted to arXiv
- License
- CC BY 4.0