An AI’s remembered lesson broke a task it could already do. Its own records held the correction.

An AI could complete a simulated task involving a vase and a desk lamp, then failed after receiving a remembered lesson. The lesson wrongly said that placing the object beside the lit lamp completed the task. The original activity record contradicted that rule. The example appears in WikiTrace, a November 10 paper from researchers at KAIST and DeepAuto.ai.

The proposed fix gives a short lesson a trail back to its evidence. Each memory links to a summary of why the earlier attempt succeeded or failed, then to the full action-and-feedback record. When a lesson lacks necessary details, conflicts with another or comes from a failed attempt, the agent can inspect those records before deciding what to do.

Across eight test suites drawn from five benchmarks, including text simulations and application-interface tasks, WikiTrace averaged 90.2 with DeepSeek-V4-Flash. The same agent averaged 79.5 without memory and 87.0 when given the same lesson summaries without the supporting records. On tasks it could previously solve, the failure rate fell from 6.5% with summaries alone to 4.5%. Average test-time cost per task rose about 17%.

The paper is under review, and these results do not establish reliability in homes or businesses. A link to a record does not automatically validate a lesson. The experiment also puts every lesson into the initial prompt, leaving large-scale memory retrieval unresolved. Its practical warning is simple: an AI’s polished takeaway can be a mistaken interpretation of an earlier outcome. Keeping the evidence makes that mistake possible to inspect and correct.

Source time: Yukyeong Lee and colleagues / arXiv, November 11, 2026, 00:46:59 Beijing time / November 10, 2026, 11:46:59 AM EST.