Shukai Liu, Jian Yang, Bo Jiang, Yizhi Li, Jinyang Guo, Xianglong Liu, Bryan Dai
The paper introduces CAT, a context management tool for software engineering agents that improves long-horizon reasoning by structuring context maintenance and enabling proactive compression of historical data.
This research addresses the challenges faced by software engineering agents working with large codebases over extended periods. Traditional methods often lead to problems like context overload and reduced reasoning ability. The authors propose a new system called CAT, which helps manage the information these agents use more effectively. By organizing context into stable task semantics, long-term memory, and short-term interactions, CAT allows agents to summarize past information efficiently, leading to better performance and reasoning over time.
