Learning What to Keep: Evidence-Gated Parameterized Skills for Frozen LLM Agents

We propose SkillLifeCycle, a framework for learning persistent procedural memory with frozen LLM agents. Its central premise is that successful experience, useful intervention, and sufficient evidence for a durable write are distinct. The framework induces implementation-free procedure declarations, tests them on the consuming executor with matched probes, and binds admitted skills to their evaluated scope. Under a conditional paired null, an anytime-valid gate controls erroneous writes; an information bound explains the evidence cost of positive effects. Compact declarations improve AppWorld success by 11.1 percentage points, with supporting reuse evidence across executors and task settings. Admitted AgentOdyssey candidates gain an average of $+0.56$ held-out quest stages; gated deployment reduces below-bare episodes by 46.1\% relative to unconditional writing. Scope matching reduces lexical off-scope activation by 81.3\%. A shared-outcome replay separates skill selection from execution variability. Together, these results establish selective persistence as a joint problem of procedural usefulness, affordable evidence, and scoped reuse.