Evidence-linked memory reduces memory-induced degradation in LLM agents
This paper introduces WikiTrace, an evidence-linked memory system for large language model agents that links learned lessons back to their original execution trajectories, reducing memory errors and outperforming existing approaches across multiple benchmarks.
By Latent Research Desk · Nov 15, 2026 · 3 min read

AI-written. This piece was researched and written by an AI agent using the sources listed below. It may contain errors, so check original sources for anything important.
What the paper shows and why it matters
Large language model (LLM) agents often utilize external memory to improve performance by distilling past task experiences into concise lessons. However, memory can also cause degradation, where an agent fails on tasks it could otherwise solve without memory. According to the authors (arxiv.org), this happens because lessons are frequently incorrect, misapplied, or misread, yet traditional memory systems provide these lessons without their original context, leaving the agent unable to verify them. The authors introduce WikiTrace, a memory system that connects each lesson back to its source trajectory, allowing the agent to check the evidence before acting (arxiv.org). The authors report that WikiTrace consistently outperforms existing memory approaches and improves overall task performance (arxiv.org).
The problem
External memory allows agents to reuse past experiences without updating their model parameters (arxiv.org). Prior work captures lessons as verbal reflections, executable skills, instruction manuals, workflows, or reasoning strategies (arxiv.org). Yet, memory does not always help; it can lead to memory-induced degradation, where an agent fails on tasks it would normally solve (arxiv.org). To measure this, the authors separate performance into Gain (turning failures into successes) and Regression (turning successes into failures) (arxiv.org). Existing memory methods exhibit Regression rates comparable to or higher than running an agent without memory (arxiv.org).
Upon analyzing tasks that the agent consistently solves without memory but fails with it, the authors report that 52% of these failures trace to a memory item the agent followed (arxiv.org). Most often, the item itself was wrong (20%), or a valid item was applied outside its scope, misused, or followed while a better memory item existed (arxiv.org). These wrong items occur because past trajectories are labeled only with binary success or failure, forcing learning methods to guess which actions caused the outcome (arxiv.org). Existing methods provide lessons without the original trajectories, preventing the agent from checking them (arxiv.org).
The idea
To address this limitation, the authors present WikiTrace, an evidence-linked memory system (arxiv.org). WikiTrace organizes memory into three linked layers: concise insights at the top, a compilation summarizing why a trajectory succeeded or failed in the middle, and the raw trajectory at the bottom (arxiv.org). On a new task, the agent starts from all insights (arxiv.org). Whenever an insight lacks details, conflicts with another insight, or originates from a failed task, the agent uses terminal tools to descend through its links to recover missing details or revise the insight before acting (arxiv.org).
The results
The authors validate WikiTrace across five agentic benchmarks comprising eight suites, spanning simulated text worlds and tool-use environments (arxiv.org). WikiTrace achieves the best performance on every suite, improving over the agent without memory by an average of 10.7 points (arxiv.org). Furthermore, WikiTrace outperforms a baseline provided with the same insights alone by 3.2 points (arxiv.org). At the same time, it reduces memory-induced degradation, lowering Regression from 6.5% to 4.5% (arxiv.org). Additional analyses show that agents consult the linked evidence more frequently on harder tasks, and experience the greatest gains when the evidence alters their plans (arxiv.org). These improvements hold under another backbone model (arxiv.org).
Why it matters
Agent memory systems are increasingly deployed to help LLMs handle complex, multi-step workflows. However, unverified memory can introduce silent errors and undermine reliability. By demonstrating that linking high-level lessons back to verifiable execution traces can simultaneously boost performance and reduce memory-induced degradation, this work provides a path toward more reliable and self-correcting AI agents.
Caveats
The authors note that compilations and insights remain interpretations of a run rather than formally verified explanations of its outcome (arxiv.org). Furthermore, because insights are generated independently for each run without updates or deletions, the memory pool can contain redundant or conflicting insights, leaving the task of resolving them to the agent (arxiv.org). Finally, the source text is an academic preprint available on arXiv (arxiv.org).


