Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
Agents can fail tasks they would otherwise solve when stored lessons are wrong—or applied incorrectly.
WikiTrace links each lesson to a summary of why its source task succeeded or failed and to the full task trajectory. Agents can inspect that evidence when a lesson lacks detail, conflicts with another, or comes from a failed task. Across five agentic benchmarks, the authors report better performance than existing memory approaches and a baseline given the same insights alone, with fewer memory-induced failures. The paper is under review.
HF Daily Papers' note
WikiTrace links each lesson to a summary of why its source task succeeded or failed and to the full task trajectory. Agents can inspect that evidence when a lesson lacks detail, conflicts with another, or comes from a failed task. Across five agentic benchmarks, the authors report better performance than existing memory approaches and a baseline given the same insights alone, with fewer memory-induced failures. The paper is under review.
HF Daily Papers' note
score 5