Google Research has developed WikiSkill, a framework that lets AI agents preserve useful lessons from previous tasks instead of starting every run from scratch. It does not continually retrain the model. It turns past execution traces into reusable instructions that can be tested, accepted or rolled back.
The system separates memory into three layers. An immutable raw layer stores tool calls and results; a wiki layer records failure patterns and successful strategies; and a skill layer contains the active procedures used by the agent. A maintainer distills new traces, a proposer suggests skill changes, and a validation gate rejects changes that reduce performance while retaining the history of what failed.
Researchers tested the approach on math, web search, spreadsheets, document questions and interactive virtual tasks. WikiSkill raised Gemini 3.5 Flash’s average score from 49.5% to 68.1% and Qwen-3.6-27B from 39.4% to 63.3% across the five benchmarks. Gains were largest in math and spreadsheet work and smaller on long-document questions. Larger models generally used evolved skills more reliably, though a smaller model with WikiSkill could sometimes match a larger baseline. These results come from controlled benchmarks, and skills did not always transfer cleanly between models. The design offers agents durable operational memory, but each proposed lesson still needs testing before deployment.