Researchers release WikiTrace, evidence-linked memory for LLM agents

A team of researchers has published WikiTrace, a memory system for LLM agents that ties every stored lesson back to the task trajectory it came from. The paper argues that agents can get worse when they reuse memory, because lessons are learned from runs labeled only as success or failure and are handed to the agent without any way to check them. WikiTrace instead keeps each insight alongside a summary of why its trajectory succeeded or failed and the raw trajectory itself. On a new task, the agent starts from the insights and follows those links, using tools, when a lesson is missing details, conflicts with another, or came from a failure. The authors report gains over existing memory methods across five agentic benchmarks, and the code is public on GitHub.