Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
Yukyeong Lee, Yumin Choi, Jinheon Baek, Sung Ju Hwang
KAIST · DeepAuto.ai
Published November 10, 2026 · on the feed November 11, 2026 · ▲ 56 · arXiv 2611.14320
The abstract
External memory enables LLM agents to improve from experience by distilling trajectories of past tasks into concise lessons for new tasks. However, memory can also lead to memory-induced degradation, where an agent fails on tasks it would solve without memory, often because it follows a lesson that is incorrect, or misapplies or misreads a valid one. This is because lessons are inferred from trajectories labeled only with task success or failure, yet are typically provided without the trajectories they came from, leaving the agent no way to check them. To address this, we introduce WikiTrace, an evidence-linked memory system designed to let agents check each lesson against its original trajectory. Specifically, motivated by the finding that a trajectory often still holds the evidence needed to catch an erroneous lesson, WikiTrace stores each lesson as a concise insight linked to a summary of why its trajectory succeeded or failed and to the raw trajectory itself. Then, on a new task, the agent starts from all insights and, whenever one lacks details needed to act, conflicts with another, or comes from a failed task, descends through its links with terminal tools to fill in missing details or correct the insight before acting. We validate WikiTrace on five agentic benchmarks, showing that it consistently outperforms existing memory approaches and a baseline given the same insights alone, while also reducing memory-induced degradation. Our code is available at https://github.com/yukyeongleee/wikitrace
The Brief
2:34A short podcast summary of the paper. Written and voiced by AI.
Transcript
Don't Just Trust Your Memory. Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents. By Yukyeong Lee and three others, from KAIST and DeepAuto AI. Published on the tenth of November.
External memory lets an LLM agent improve from experience. Trajectories of past tasks are distilled into concise lessons, and those lessons are handed to the agent on new tasks. Memory can also hurt. The authors call it memory-induced degradation: the agent fails on a task it would have solved with no memory at all. Often the cause is a lesson that is incorrect, or a valid lesson that the agent misapplies or misreads.
The reason lies in how lessons are made. They are inferred from trajectories labeled only with task success or failure, which is thin ground for a lesson to stand on. Yet they are typically provided without the trajectories they came from, so the agent has no way to check them. A lesson arrives as a conclusion with its reasoning stripped away, and the agent can only take it or leave it.
WikiTrace is an evidence-linked memory system, designed so that an agent can check each lesson against the trajectory it came from. It is motivated by a finding: a trajectory often still holds the evidence needed to catch an erroneous lesson. The lesson can be wrong while the record behind it remains intact.
Each lesson is stored in layers. On top sits a concise insight. That insight is linked to a summary of why its trajectory succeeded or failed, and through that to the raw trajectory itself. On a new task, the agent starts from all the insights. Three situations send it deeper: an insight lacks details needed to act, an insight conflicts with another one, or an insight comes from a failed task. In any of those cases, the agent descends through the links with terminal tools, to fill in what is missing or to correct the insight, before it acts. The three triggers line up with the failures named earlier. A lesson too vague to act on invites misapplication. Two lessons in conflict mean at least one may be wrong. And a lesson drawn from a failed task rests on the weakest kind of label.
WikiTrace is validated on five agentic benchmarks. It consistently outperforms existing memory approaches. It also outperforms a baseline given the same insights alone, which is the comparison that isolates the links: the same lessons, with and without the evidence behind them. And it reduces memory-induced degradation, the failure the system was built to address.
Instead of trusting a lesson or throwing it away, WikiTrace keeps the path back to where the lesson came from. Most of the time the agent reads only the short insight. When something about it looks incomplete, contradictory, or drawn from a failure, the agent has a way to go and look, and on five benchmarks that habit of checking is what separates it from memory that has to be taken on trust.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Memory should make an agent better. Sometimes it makes it worse.
Your LLM agent keeps short notes distilled from its past tasks and reads them before each new task. Usually that helps it avoid repeating mistakes.
The road ahead
- 1 · The problem ~6 min Notes distilled from past runs can make an agent fail tasks it would solve without them, and it has no way to check them.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Keep every lesson linked down to the evidence it came from, and let the agent read that evidence when a lesson looks doubtful.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Build three linked layers from past runs once, then let the agent read all insights and descend through links when needed.
- 6 · The experiment ~6 min You’ll guess: Against a baseline given the same insights without links to evidence, how did WikiTrace do across five agentic benchmarks?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.