Procedural memory—a continuously updated library of natural-language design skills—enables frozen frontier models to improve at complex agentic tasks by learning from execution failures without model retraining or human annotation.
This paper shows how a frozen AI model can continuously improve at graphic design by building and refining a library of reusable design procedures from real user projects. Without updating the model's weights or using human labels, the system learns 139 design skills from 1,406 real briefs, improving success rates from 73% to 99% by accumulating new procedures and fixing failed ones.