Google's WikiSkill gives AI agents a persistent memory of past mistakes to sharpen future performance

Google Research introduced WikiSkill, a framework that pairs AI agents with a persistent knowledge base of past failures and successes to improve performance over time. The system distills execution traces into a wiki layer and reusable "Agent Skills," with a gating mechanism that rolls back unhelpful changes. Inspired by Andrej Karpathy's "LLM Wiki" idea, WikiSkill outperformed prior skill evolution methods, boosting Gemini-3.5-Flash from 49.5 to 68.1 percent on average across five benchmarks.