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

What it claims

A continual-adaptation framework where a frozen frontier model drives professional design software through 230+ tools, while an external, natural-language "procedural memory" of reusable design skills grows and self-revises from experience — no weight updates, no human labels. Over five rounds on 1,406 real user briefs and 1,869 auto-graded trajectories, the skill bank grows from 76 documentation-derived skills to 139, and GenEval2 execution success on Claude-Sonnet-4 rises from 72.7% to 99.3%.

How it works

The memory changes along two axes, and the ablation is the point of the paper:

• Widening: acquiring new procedures for recurring subtasks the current bank does not cover.
• Deepening: revising existing procedures against their own successful and failed executions.
• A matched replay gate admits a proposed change only if it repairs failures without regressing previously observed successes — the safeguard that makes editing under noisy, oracle-free feedback safe.

The framing motivating this: professional graphic design is a long-horizon agentic task producing structured editable artifacts, with no reliable programmatic oracle for correctness, so the system must adapt under unverifiable feedback.

flowchart TD
  E[Graded trajectories] --> W[Widen: new skills]
  E --> D[Deepen: revise skills]
  W --> G{Matched replay gate}
  D --> G
  G -->|repairs w/o regression| M[Procedural memory bank]
  G -->|regresses successes| X[Rejected]
  M --> A[Frozen model + 230+ tools]
  A --> E

Evidence

• GenEval2 execution success, Claude-Sonnet-4: 72.7% → 99.3% (+11.99 points in generation quality).
• Win rates vs. a no-skill agent across four specialized design benchmarks: 61.8% on Claude-Sonnet-4, 67.6% on Claude-Opus-4.6.
• Ablation on 200 held-out briefs from a user-traffic benchmark: widening-alone 49.4%, deepening-alone 48.6%, combination 58.5% (p = 0.025) — the two mechanisms are complementary, not redundant.

Limits

• This is the abstract only; the fetched page gives no method internals beyond the above, no tool list, no code or data availability, and no licence.
• All gains come from a frozen base model with an external memory — no fine-tuning, so results depend on the underlying frontier model's capability (Sonnet-4 / Opus-4.6 here).
• Grading is automatic ("automatically graded trajectories") for a task the authors themselves say has no reliable oracle; the win-rate metric is comparative against a no-skill baseline, not an absolute quality measure. Confirm how "GenEval2" and the four benchmarks are defined before trusting the headline number.