SPEM introduces structure-aware procedural memory for LLM agents

SPEM is a memory framework for LLM agents that uses Skill-DAGs to store reusable skill compositions and dependencies. By maintaining both atomic and procedural memory, agents can efficiently select and execute complex skill sequences via a Graph Neural Network.