Agent memory is usually sold as pure upside: the agent writes down what it learned and does better next time. A preprint from KAIST and DeepAuto.ai puts a number on the downside. When the authors looked at tasks an agent reliably solved without memory but failed with it, 52% of those failures traced back to a stored lesson the agent had followed, most often a wrong one.
The cause is simple. A past run is labelled only as a success or a failure, so whoever writes the lesson has to guess which of many actions mattered. WikiTrace keeps the evidence. Each lesson is a short insight that links to a compiled diagnosis of why its run went well or badly, and from there to the raw trajectory. The agent reads every insight at the start of a task, and when one lacks detail, contradicts another or came from a failed run, it follows the links with ordinary terminal tools before acting.
Tested on eight suites across five benchmarks, from household and science text worlds to AppWorld and a retail customer-service test, WikiTrace scored best on every suite. It beat the agent without memory by 10.7 points on average and a version given the same insights without links by 3.2 points. The share of previously solved tasks that memory broke fell from 6.5% to 4.5%. At test time it cost 1.17 times as much per task as the insights alone. A second backbone, GLM-5.3-flash, showed the same pattern on three suites.
For anyone running agents with Mem0, Letta or claude-mem, the useful idea is measurement as much as method. The authors split results into gain, tasks memory fixes, and regression, tasks it breaks, because an average hides the second. Their figures show existing memory methods breaking tasks at a rate close to, or above, simply rerunning the agent with no memory at all.
The caveats are the authors' own. A link lets the agent check a lesson but does not prove it right, and a diagnosis drawn from a pass or fail label can still be wrong. Every insight goes into the first message, so the prompt grows with the memory, and finding the right lessons in a store too big to load is left unsolved. The causes of degradation were labelled by an LLM judge reading the agent's reasoning. This is a preprint, not yet peer reviewed.