KAIST's WikiTrace lets agents check stored lessons against their source runs
KAIST researchers describe memory-induced degradation and introduce WikiTrace, a memory system that links each lesson to the run it came from.
KAIST researchers report that external memory can make an LLM agent worse at tasks it already handles. They propose WikiTrace, a memory design that keeps each stored lesson tied to the run it was drawn from, so the agent can check it before use. The authors call the failure memory-induced degradation: an agent fails tasks it would solve without memory, often because it follows a lesson that is wrong or misreads a valid one. According to the paper, lessons are inferred from trajectories labeled only as success or failure. They are usually handed to the agent without those trajectories, so the agent has no way to verify them.
WikiTrace works in two layers. A compiler turns each past run into a diagnosis of why it succeeded or failed, plus a set of short insights. Each insight links back to that diagnosis and to the raw trajectory. On a new task, the agent starts from the full set of insights. When one lacks the details needed to act, conflicts with another, or comes from a failed run, the agent follows its links with terminal tools to fill in or correct it before acting. A trajectory often still contains the evidence needed to catch a bad lesson, so the lesson does not have to be trusted on its face.
The authors evaluated on ALFWorld, ScienceWorld, PlanCraft, AppWorld and τ³-bench Retail. WikiTrace scores 10.7 points above an agent with no memory on average, and 3.2 points above a baseline given the same insights without the links. That 3.2-point comparison isolates the links from the insights themselves. The failure rate on tasks the agent could solve without memory falls from 6.5% to 4.5%.
The curator summary in the paper adds that agents consult the linked evidence more on harder tasks, where the advantage is largest. The paper is a preprint submitted to the ACL Rolling Review October 2026 cycle and is still under review. The authors say their code is available on GitHub.
What changed
WikiTrace links each stored agent lesson to the run it came from so the agent can check the lesson against its evidence before acting, instead of trusting it directly.
Type: Research · Applies to: ALFWorld, ScienceWorld, PlanCraft, AppWorld, τ³-bench Retail
Why it matters
Agent memory can make agents worse at tasks they already solve, so keeping evidence beside each stored lesson may matter for production agents that learn from experience.
Original source
Where we saw it
DAIR.AI (Elvis Saravia)
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works