WikiTrace links each agent memory lesson back to the trajectory it came from
A KAIST AI paper argues that agent memory can make agents fail tasks they would otherwise solve, and proposes WikiTrace, which lets the agent check each stored lesson against its original trajectory.
The problem
External memory is a common way to let LLM agents improve from experience. Past task trajectories are distilled into short lessons, and those lessons are handed to the agent on new tasks.
A paper from KAIST AI, published on arXiv on 2026-11-10 and listed on Hugging Face daily papers on 2026-11-09, focuses on a downside the authors call memory-induced degradation. This is the case where an agent fails a task it would have solved without memory. According to the abstract, it often happens because the agent follows a lesson that is incorrect, or misapplies or misreads a valid one.
The authors trace this to how lessons are made. Lessons are inferred from trajectories that are labeled only with task success or failure. They are then usually given to the agent without the trajectories they came from, so the agent has no way to check them.
How WikiTrace works
The proposed system, WikiTrace, is an evidence-linked memory. It is motivated by a finding reported in the abstract: a trajectory often still holds the evidence needed to catch an erroneous lesson.
WikiTrace stores memory in three linked layers:
- a concise insight, which is the lesson itself;
- a summary of why the source trajectory succeeded or failed;
- the raw trajectory.
On a new task, the agent starts with all the insights. It does not always trust them. Whenever an insight lacks details needed to act, conflicts with another insight, or comes from a failed task, the agent descends through its links using terminal tools. It uses what it finds to fill in missing details or correct the insight before acting.
In this design, the short insights are always in front of the agent, and the summaries and raw trajectories are consulted only when one of those three conditions applies.
Reported results
The authors validate WikiTrace on five agentic benchmarks. According to the abstract, it consistently outperforms existing memory approaches and also beats a baseline that is given the same insights alone, without the links. It also reduces memory-induced degradation.
The comparison with the insights-only baseline is the most direct test of the idea, since it separates the value of the links from the value of the lessons themselves.
What the abstract leaves open
The abstract does not name the five benchmarks or the existing memory approaches, and it gives no numbers for either the accuracy gains or the reduction in memory-induced degradation. It also does not report how often agents descend into trajectories, or what that costs in extra tokens and tool calls. The authors are Yukyeong Lee, Yumin Choi, Jinheon Baek and Sung Ju Hwang, and the code is available in a public GitHub repository.