Persistent LLM agents currently possess episodic and semantic memory, but lack procedural memory—know-how, internalized procedures. A human who has resolved ten DNS failures has internalized a diagnostic procedure; an LLM agent has ten separate memories it must rediscover each session. We formalize procedural memory as consolidated action patterns automatically extracted from repeated experiences, and propose a concrete architecture with extraction, storage, and incremental refinement. Also available in French: Mémoire procédurale pour agents LLM persistants : le savoir-faire comme composant manquant