A production service called Zep Agent Skills provides procedural memory to deployed LLM agents using only verified task attempts as evidence. A compiler with Reflector and Curator roles transforms these verified traces into versioned Skills, governed by admission gates. The system uses these Skills, which are open standard, for agents to retrieve knowledge through lexical search. Compiling skills from a teacher model significantly increases the pass rate of weaker executor models on tasks. This approach demonstrates substantial gains in performance without requiring continuous fine-tuning.