WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution
Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan and Tu Vu (Google) introduce WikiSkill, which co-evolves executable agent skills with a persistent wiki-shaped knowledge base so that the insights behind each skill revision survive past the optimization run that produced them.
Ask this paper
Three-layer separation: Raw execution experience, accumulated knowledge (the wiki), and executable skills are kept as distinct artifacts. Most skill-evolution work collapses the middle layer into optimization history, which is why insights stop compounding after a few iterations.
Ablations pin the gain on the wiki: Removing persistent knowledge accumulation degrades skill evolution, so the win is the durable memory substrate rather than better skill-writing prompts.
Skills substitute for scale: Smaller models equipped with evolved skills outperform substantially larger models without them, and larger models still benefit more from skills, so skill evolution complements rather than replaces scaling.
Cross-model transfer: Skills evolved by one model transfer across models and families, and skills evolved by another model can beat self-evolved ones, which argues for shared skill registries over per-model libraries.
Why it matters: This is the cleanest statement yet that agent memory should be wiki-shaped: consolidated, addressable, and separate from both the trace log and the executable artifact.
Abstract
Agent skills package specialized knowledge and workflows into reusable resources that extend AI agent capabilities. Recent work automatically discovers such skills from agent experience, which enables agents to progressively adapt through interaction. However, the insights that guide skill development typically remain scattered across optimization histories, limiting their systematic reuse across iterations. We introduce WikiSkill, a framework that co-evolves agent skills with a persistent knowledge base (wiki). At a high level, WikiSkill separates raw execution experience, accumulated knowledge, and executable skills, while continuously consolidating experience into the wiki, which subsequent skill updates can build on. Across diverse benchmarks and models, WikiSkill consistently outperforms state-of-the-art skill-evolution methods and improves over no-skill baselines in most model-benchmark settings. We find that skill evolution complements model scaling: larger models generally benefit more from evolved skills, while smaller models with skills can outperform substantially larger models without them. We also find that evolved skills transfer effectively across models and model families, and skills evolved by other models can outperform self-evolved skills. Finally, our ablation studies confirm that persistent knowledge accumulation in the wiki is critical for effective skill evolution. These results demonstrate the benefits of systematically accumulating and refining agent experience for developing reusable and transferable skills.