Bidirectional Coevolution of Memory and Skills for Lifelong LLM Agents

LLM agents that operate over long periods can preserve completed interactions in external memory and compile recurring episodic evidence into reusable skills. Prior work has advanced memory updated by feedback and skill libraries that evolve during use. Recent systems begin to maintain both. We introduce MemCryst, a framework that learns from deployment experience by updating external memory and skills while keeping the language model frozen. Each completed interaction writes episodic memory and may contribute typed semantic facts. Recurring episode clusters provide evidence for creating and revising audited conditional skills. Retrieved skills shape later behavior and therefore the episodes written next. Outcomes from joint fact and skill use update their retrieval influence through task success statistics without rewriting record text. We evaluate ten task settings spanning embodied control, reading long documents, conversation, mathematics, and code. Persistent streams are scored prequentially, so every initial error counts. MemCryst leads seven of ten comparisons, and the deployed ALFWorld system raises success from $51.5$ to $63.17$. A matched $2 \times 2$ intervention over 600 tasks yields an observed interaction of $+5.3$ percentage points between memory and skill feedback.