Feed
WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution
An AI agent keeps a running notebook of lessons, then writes better how-to guides from it.
47.4%
Qwen-3.5-9B with WikiSkill, avg accuracy
A 9B model with skills beats a 27B model without them
How it works
Lessons get scattered
Methods that auto-write agent skills lose their insights across optimization histories, so each new edit starts half-blind.
Keep a persistent wiki
A maintainer agent turns execution traces into wiki pages of failure patterns. A proposer edits skills from the wiki, and validation gates each edit.
Better, transferable skills
Highest average accuracy for all five models. Skills also transfer across model families, sometimes beating self-evolved ones.
From the paper