Don't trust your memory alone: evidence-linked memory reduces memory-induced degradation in LLM agents

Researchers propose an evidence-linked memory system, WikiTrace, which links each lesson distilled by an agent to its source trajectory, letting the agent verify and correct its memory before acting, thereby reducing memory-induced degradation. The paper points out that although external memory can let agents improve from experience, it can also cause memory-induced degradation, that is, the agent, by following wrong lessons or misusing or misreading correct lessons, performs poorly on tasks it could have completed without memory, and lessons usually carry no source trajectory and cannot be checked. WikiTrace stores each lesson as a brief insight, linked to a summary of the success or failure reasons of its trajectory and to the original trajectory; when facing a new task, it starts from all insights, and when an insight lacks action details, conflicts with other insights, or comes from a failed task, it uses terminal tools to drill down along the links, completing missing details or correcting the insight before acting. On five agent benchmarks, WikiTrace consistently outperforms existing memory methods and a baseline that provides only the same insights, and the code is open-sourced at https://github.com/yukyeongleee/wikitrace。