SPEM: Structure-Aware Procedural Memory for Efficient Skill Use in LLM Agents
SPEM is a lightweight structure-aware procedural memory framework for efficient skill use in LLM-based agents and achieves the best overall effectiveness--efficiency trade-off, reaching 87.1% ALFWorld success and 38.38% SkillsBench reward.
Ran Zhou, Ze-Hui Kong, Fan-Yu Lin et al.
· 0 citations