← Back to digest
AI Agents

WikiSkill: compiling agent experience into persistent knowledge

What's this about?

Agent skills package specialised knowledge and workflows into reusable resources. Recent work discovers such skills automatically from agent experience. WikiSkill takes the next step: compiling that experience into persistent knowledge that survives and evolves across runs, rather than being rediscovered each time.

What is the problem being solved?

Agents are amnesiac by default. An agent that spends forty minutes working out how your build system handles a particular edge case throws that away when the session ends, and the next agent pays the same forty minutes. Persistent skill compilation is the attempt to make the second run cheap — which sounds obvious, and is hard, because you have to decide what generalises and what was specific to that one situation.

Why does this matter to you?

  1. It is the research version of what skills and memory files do informally in tools you already use.
  2. The interesting question it raises is quality control: accumulated skills can encode mistakes just as persistently as insights.
  3. Pairs with JIT-Agent — both are arguing the durable value is in the scaffolding, not the model.
← SWE Refactor BenchAutonomous mathematical discovery in "the Station" →