Separating raw experience, a persistent wiki, and executable skills beats every skill-evolution method that keeps them in one pile

arXiv 2608.27454·medium signalXBlueskyLinkedInCopy link

WikiSkill co-evolves an agent's skills with a persistent knowledge base, consolidating execution experience into the wiki so later skill updates build on it instead of re-deriving insight from scattered optimization histories. It beats state-of-the-art skill evolution across benchmarks and models, and the ablations pin the gain on the wiki itself. Two transfer results matter for anyone maintaining a skill directory: skills move across model families, and skills evolved by another model can beat self-evolved ones, so a small model with good skills can outrun a substantially larger one without them.