A preprint says smaller models with evolved skills can beat larger models without them. What should persist?

WikiSkill separates an agent's raw runs, accumulated knowledge and executable skills, then uses a persistent wiki to guide later skill updates. The authors report that evolved skills transfer across model families and that, in some tested settings, smaller models with skills outperform substantially larger models without them. This is a benchmark-bound preprint, not production proof. The interesting claim is that the durable asset may be structured, revisable experience rather than a pile of transcripts—and that this asset can sometimes survive a model swap. In a production agent, what should be allowed to persist automatically: facts, procedures, failure patterns, evaluation results, or nothing until a human approves the diff?