WikiSkill
A Google Research method that gives an AI agent a growing written record of its own successes and failures, which it reads before starting new work.
WikiSkill is a framework described by Google Research and Virginia Tech in August 2026. It addresses a simple weakness in AI agents: they forget. An agent finishes a task, and everything it learned along the way disappears. WikiSkill keeps three layers instead. A raw record of every step taken, a wiki of distilled lessons about what failed and what worked, and a set of active instructions the agent actually follows, which can be rolled back if an update makes things worse.
It is worth being clear about what is learning here. The model itself does not change at all. What improves is the written guidance it reads before it starts. That makes this closer to external memory and automated prompt refinement than to a model that keeps training after release, and it is a good example of how much can be gained without touching the weights.