KAIST WikiTrace Tackles Agent Memory-Induced Degradation
The counterintuitive finding is that memory can be a liability, not just an asset, for agents — a failure mode most memory frameworks do not test for. Builders of long-running agent systems should treat provenance and traceability of stored lessons as a design requirement rather than an afterthought. The benchmark spread across five environments makes the results easy to compare against your own memory stack.
KAIST researchers identify a failure mode they call memory-induced degradation: an agent's accumulated memory can cause it to fail tasks it would otherwise solve. The problem arises because lessons are inferred from runs labeled only as success or failure and saved without the underlying run context, so a wrong lesson cannot be verified by the agent.
The team introduces WikiTrace, which stores each lesson as a short insight linked to a summary of why the run succeeded or failed, plus the raw trajectory. When an insight is vague, conflicts with another insight, or originates from a failed run, the agent follows those links using terminal tools before acting on it.
Across ALFWorld, ScienceWorld, PlanCraft, AppWorld, and tau3-bench retail, WikiTrace scores 10.7 points above a memory-less agent on average and 3.2 points above using the same insights alone. The failure rate on previously solvable tasks drops from 6.5% to 4.5%.
Interesting insights about agent memory.
Might be worth a read if you are building memory systems for your agents.
Can an agent's memory make it fail tasks it would solve without memory?
KAIST researchers find that it can, and they call it memory-induced degradation.
The cause is how lessons are stored. They infer lessons from runs labeled only as success or failure, then save them without the run itself. When a lesson is wrong, the agent has no way to verify it.
The authors introduce WikiTrace, which stores each lesson as a short insight linked to a summary of why the run succeeded or failed and to the raw trajectory. When an insight is vague, conflicts with another, or comes from a failed run, the agent follows the links with terminal tools before acting.
Across ALFWorld, ScienceWorld, PlanCraft, AppWorld, and tau3-bench retail, it scores 10.7 points above the agent without memory on average and 3.2 points above the same insights alone. The failure rate for previously solvable tasks drops from 6.5% to 4.5%.
https://t.co/3wU1elq35o
Chat with Paper: https://t.co/b182EkAvLh
This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include KAIST, WikiTrace, agents, AI safety.
Source attribution: curated from DAIR.AI (@dair_ai).
