Google's WikiSkill gives AI agents a persistent memory of past mistakes to sharpen future performance
decoder ·
Google Research has introduced WikiSkill, a framework that gives AI agents a persistent, wiki-like knowledge base to learn from past mistakes and successes. Instead of resetting after each run, the system stores execution traces, distills them into structured insights, and generates reusable 'Agent Skills' that guide future behavior without retraining the model. The framework separates experience, knowledge, and action into three layers: a raw trace layer, an immutable wiki layer, and a skill layer that can be rolled back if updates hurt performance. A cycle involving a Wiki Maintainer, Skill Proposer, and gating mechanism continuously updates and validates skills. In tests across five benchmarks, WikiSkill consistently outperformed other skill evolution methods, improving Gemini-3.5-Flash from 49.5% to 68.1% average accuracy and Qwen-27B from 39.4% to 63.3%. Larger models benefited more, but smaller models with WikiSkill could match larger models without it, and skills sometimes transferred between models.
decoder