Google's WikiSkill gives AI agents a persistent memory of past mistakes to sharpen future performance

Details

Google Research has introduced WikiSkill, a framework that gives AI agents structured, persistent memory across task runs. Inspired by Andrej Karpathy's concept of an 'LLM Wiki,' the system organizes agent knowledge into three tiers: a Raw Layer of immutable execution traces, a Wiki Layer of distilled failure patterns and successful strategies that never resets, and a Skill Layer of active procedural instructions that can be rolled back if they hurt performance. The four-step cycle works as follows: an inference agent executes tasks and logs traces; a Wiki Maintainer analyzes those traces and writes findings into the persistent wiki; a Skill Proposer suggests targeted skill updates based on the wiki; and a gating mechanism validates each update on a held-out set before applying it. Crucially, failed proposals aren't discarded—the wiki documents what was tried and why it didn't work, so future iterations avoid re-testing dead ends. The researchers are clear that agents aren't truly learning continuously—that remains an unsolved problem. Instead, they write better instructions for themselves after each run and retrieve those instructions next time. The practical payoff is real: smaller models with WikiSkill can reach performance levels of larger models without it, and larger models scale further with the framework. This sits in the growing space of inference-time scaffolding augmentation, where the model itself doesn't change but its runtime environment accumulates structured experience—reducing repeated failures on known problem patterns without touching weights.