WikiSkill is a framework for evolving AI by adding a persistent knowledge base (wiki) that separates raw execution traces, accumulated knowledge, and executable skills. Each iteration runs the agent on tasks, consolidates successful and failed patterns into the wiki, and then proposes skill updates informed by that stored knowledge. This allows skill development to build on previously captured insights rather than starting from scratch.
Tested across multiple benchmarks and language models, WikiSkill consistently outperformed prior skill-evolution methods (EvoSkill, SkillOpt, Trace2Skill) and often beat the no-skill baseline. The benefits scaled with model size: larger models gained more from evolved skills, but smaller models with WikiSkill skills could beat larger models without any skills. Skills transferred effectively between different models and model families, and in some cases skills evolved by one model outperformed those a target model evolved for itself.
What makes this interesting is the demonstration that systematically accumulating agent experience into a structured knowledge layer can make more efficient and yield skills that are more general and reusable. The paper also shows that preserving the wiki across iterations is critical—ablation studies confirmed that removing the persistent knowledge base degrades performance significantly.