Research Editor selected
Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
The paper identifies memory-induced degradation in LLM agents, where retrieved lessons cause failures on tasks the agent could otherwise solve. It introduces WikiTrace, an evidence-linked memory system that stores each lesson with a summary of its source trajectory's success or failure and the raw trajectory itself. Agents can descend through these links to verify, fill in, or correct lessons before acting. Experiments on five agentic benchmarks show WikiTrace outperforms existing memory approaches and a baseline given the same insights alone while reducing memory-induced degradation.
Why it matters: It offers a concrete memory-design method for improving LLM agent reliability and mitigating failures caused by unverified past lessons.
LLM agentsmemory systemsagent reliability