Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design

Designer-RSI. Evolving Procedural Memory from User Traffic for Agentic Graphic Design. By Hongyang Du and three others, from Adobe and Brown University. Published on the eighteenth of September. No code repository is linked to it.

Professional graphic design is a long-horizon agentic task. Structured, editable artifacts emerge from many interdependent actions, and the outcomes admit no reliable programmatic oracle. There is no test to run that says a design is right.

The authors propose a framework of continual adaptation. A frozen frontier model operates professional design software through more than two hundred and thirty tools. Next to it sits an external procedural memory of natural- language skills, which accumulates reusable design procedures from experience and refines them. The model does not change. The memory does. Everything the agent learns over time is written down as procedures in plain language, where it can be read, revised, or removed.

The memory grows in two ways. It widens by acquiring procedures for subtasks that recur but are not yet covered. It deepens by revising existing procedures against their own successful and failed executions. A matched replay gate governs both, admitting only changes that repair failures without regressing successes already observed. That gate is what lets the memory learn from noisy feedback without sliding backward. Widening adds breadth, deepening adds reliability, and the gate makes sure neither costs the agent something it could already do.

Five rounds ran over one thousand four hundred and six real user briefs and one thousand eight hundred and sixty-nine automatically graded trajectories, with no weight updates and no human labels. The bank grew from seventy-six skills derived from documentation to one hundred and thirty-nine.

GenEval2 execution success on Claude-Sonnet-4 rose from seventy-two point seven percent to ninety-nine point three percent, with a gain of eleven point nine nine points in generation quality. Against the agent with no skills, across four specialized design benchmarks, win rates reached sixty-one point eight percent on Claude-Sonnet-4 and sixty-seven point six percent on Claude- Opus-4.6.

The two growth mechanisms also work in combination. On two hundred held-out briefs from the user-traffic benchmark, widening alone reaches a forty-nine point four percent win rate over the no-skill agent, and deepening alone forty-eight point six percent. Together they reach fifty-eight point five percent, with p equal to zero point zero two five.

No weights move and no one labels anything. The model stays frozen while what it knows how to do lives in a memory that grows wider and deeper from real use, which the authors offer as a practical route to continual adaptation under feedback that is noisy and cannot be verified.