WikiTrace Links LLM-Agent Lessons to Source Runs, Hits 90.2%

External memory can make LLM agents worse at tasks they would otherwise solve, a new Hugging Face paper argues, and the fix it proposes is keeping every lesson linked to the raw trajectory it came from. On eight test suites spanning ALFWorld, ScienceWorld, PlanCraft, AppWorld, and τ³-bench retail, the system, WikiTrace, averages 90.2% versus 79.5% with no memory, and cuts what the authors call 'memory-induced degradation' from 6.5% to 4.5%.

The paper defines that failure mode as one 'where an agent fails on tasks it would solve without memory, often because it follows a lesson that is incorrect, or misapplies or misreads a valid one.' The diagnosis the authors offer: 'lessons are inferred from trajectories labeled only with task success or failure, yet are typically provided without the trajectories they came from, leaving the agent no way to check them.'

WikiTrace stores each lesson as a concise insight linked to a summary of why its trajectory succeeded or failed and to the raw trajectory itself. When an insight lacks details to act on, conflicts with another, or comes from a failed run, the agent 'descends' through those links with terminal tools to read the original run. The claim is that 'a trajectory often still holds the evidence needed to catch an erroneous lesson.'

Descent happens 22% of the time on easy tasks and 44% on hard ones, at 1.17x the per-task cost of injecting the same insights without evidence access. Experiments run on DeepSeek-v4-flash, one of 56 DeepSeek-tagged alerts we have tracked in the last 90 days, with code published at github.com/yukyeongleee/wikitrace.