Most People Think Improving an AI Agent Means Retraining the Model. That's One Lever. There Are Four.

Most people think improving an AI agent means retraining the model.

That's one lever. There are four.

A new paper from the Memento team just demonstrated a system where agents improve themselves continuously — without any model updates. The agent rewrites its own skill library based on what worked and what didn't. Accuracy on multi-step reasoning tasks went from 65% to 92% in three cycles. On Humanity's Last Exam, performance more than doubled. Starting from five basic skills, the library grew to 235. No fine-tuning. No human intervention.

This is impressive. But the bigger insight isn't the paper itself. It's what the paper reveals about the full optimization surface that most enterprises aren't using.

There are at least four independent levers to improve an agent's performance:

Model optimization. Fine-tune the underlying SLM on your domain data. The SLM Flywheel. This changes what the model knows. Checkr did it — 5× cheaper, 30× faster, 90%+ accuracy.

Skill optimization. Evolve the agent's behavioral playbook — the workflows, multi-step procedures, and decision trees it follows. This is what Memento-Skills demonstrates. It changes how the agent acts, without changing what it knows.

Prompt optimization. Refine the instructions, system prompts, and few-shot examples. A/B test prompts against production outcomes. This changes how the agent interprets tasks. The cheapest lever, often the most underused.

Tool optimization. Improve which tools the agent selects, when it uses them, and how it chains them. Better tool routing, smarter fallback logic, tighter integration. This changes what the agent can do.

Most enterprise teams pull one lever — usually the model — and leave the other three untouched. That's like optimizing your engine but never changing the tires, the route, or the driver's technique.

The compound effect is multiplicative, not additive. A 20% improvement on each lever doesn't give you 80% total improvement. It gives you 2× or more. Because each lever amplifies the others: a better model makes skill optimization more effective, which makes better use of tools, which generates better training data for the next model cycle.

This is the optimization flywheel. Not just a model flywheel — a full-stack agent improvement loop that touches every layer simultaneously.

The teams winning at enterprise AI aren't the ones with the best model. They're the ones pulling all four levers at once.

Paper: Memento-Skills: Evolving LLM Agents Through Self-Improving Skill Memory