SimSkill is an LLM agent that treats the SUMO traffic simulator as an environment to explore. It repeatedly finds capability gaps, generates tasks to fill them, solves those tasks, verifies the solutions with an action-critic loop, and stores what worked in three kinds of memory: episodic records, procedural skills, and semantic knowledge.
On two held-out benchmarks across three backbone LLMs, independently verified results improved by up to 25 percentage points. Ablations show procedural and semantic memory contribute complementary gains. The benefit is not universal: memory helped certain backbones more than others and did not always lower inference cost.
The interesting part is the design: all control logic, operating principles, and accumulated knowledge are in natural language, with an LLM combining them with executable code. That makes accumulated capability inspectable and reusable rather than hidden in weights.