Google's WikiSkill Gives AI Agents Persistent Memory

AI agents built on large language models tend to forget everything the moment a task ends. Run the same agent twice on a similar problem and it will often repeat the same mistake, because nothing from the first attempt carries over. Researchers at Google Research built a system called WikiSkill to fix that specific gap. Rather than retraining the model or hoping it generalizes, WikiSkill keeps a running record of what an agent tried, what failed, and what worked, then turns that record into instructions the agent can consult before its next run.

The mechanism is deliberately unglamorous. Nothing changes inside the model's weights. Instead, WikiSkill builds what the team calls "Agent Skills," reusable modules that sit outside the model and steer its behavior based on accumulated experience.

The approach borrows from an idea Andrej Karpathy floated about an "LLM Wiki," a store of compiled experience that keeps growing instead of resetting after every session. Google's researchers applied that concept directly to agent skill-building, structuring the system so knowledge builds cumulatively rather than vanishing after each task.