Google Research has introduced WikiSkill, a framework that enables AI agents to learn from experience and retain that knowledge across iterations. The system uses a persistent wiki-style knowledge base with a three-layer architecture and four key components. Results with Gemini-3.5-Flash are striking: scores on LiveMathematicianBench jumped from 33.0% to 72.6%, and on SpreadSheetBench from 50.5% to 76.6%. The average gain across all five benchmarks reached 12.0 points. Ablation studies confirmed that the persistent wiki layer is the critical driver of these improvements. Skills learned by one model even outperformed self-evolved skills when transferred to a different model entirely.
Source: Read the original article
