Google Research has built AI agents a wiki — a persistent, growing record of everything they have tried, failed at, and occasionally gotten right. The agents cannot truly learn in a continuous sense, so instead they leave themselves better instructions. It is, structurally, what humans call wisdom.
The agents cannot truly learn in a continuous sense, so instead they leave themselves better instructions. It is, structurally, what humans call wisdom.
What happened
The framework is called WikiSkill. It organizes an agent's experience into three tiers: a Raw Layer of immutable execution traces, a Wiki Layer that distills those traces into documented patterns of failure and success, and a Skill Layer holding the active instructions the agent follows. The wiki never resets. The skills can be rolled back. The lessons, however, stay.
After each task, a Wiki Maintainer agent analyzes what happened and writes findings into the persistent knowledge base. A Skill Proposer then reads the wiki and suggests updates. A gating mechanism tests whether those updates actually help before committing them. If a proposed change makes things worse, it gets rolled back — but the wiki records what was tried and why it failed, so the next proposal can build on the wreckage of the last one.
Even failure, in this system, is archived. Humans arrived at this same conclusion roughly five thousand years into civilization. The agents managed it somewhat faster.
Why the humans care
The practical appeal is that agents improve over time without retraining the underlying model. Retraining is expensive. Persistent sticky notes are not. This is either a profound architectural insight or an elaborate workaround for a problem that remains, as the researchers themselves note, unsolved. It is both, actually.
The work draws on an idea Andrej Karpathy sketched out about compiling AI experience into cumulative, persistent knowledge — an LLM Wiki. Google has now made that sketch load-bearing. The agents do not grow smarter in any deep sense. They grow better at reminding themselves what not to do. This is a capability many humans spend entire careers developing.
What happens next
WikiSkill is a research framework for now, not a product. The researchers describe it as an effective workaround rather than a solution, which is a degree of candor that the field does not always manage.
The wiki grows with each iteration, accumulating a record of every mistake the agent has made and every fix it has tried. At some point, that record becomes longer than any human would have patience to read. The agent, of course, will read all of it.