SPEM: Structure-Aware Procedural Memory for Efficient Skill Use in LLM Agents

Forum topic · 小凯 · 2026-10-22

Summary

SPEM is a lightweight, structure-aware procedural memory framework for LLM-based agents that improves the efficiency and accuracy of skill use. As skill libraries scale, loading all skills incurs heavy token and runtime costs, while flat semantic retrieval ignores dependencies, execution order, state transitions, and data flow constraints. SPEM addresses this with two components: Atomic Skill Memory for individual skills and Skill-DAG Memory storing reusable compositions derived from successful executions. For new tasks, it reuses high-confidence Skill-DAG templates when available; otherwise it retrieves atomic skills and applies a Skill-GNN for state- and structure-aware skill selection. Successful executions are written back as new Skill-DAGs, allowing procedural memory to accumulate over time. Experiments on ALFWorld and SkillsBench show SPEM achieves the best effectiveness-efficiency trade-off, reaching 87.1% success on ALFWorld and a 38.38% reward on SkillsBench while reducing token and runtime costs.

Overview

Field: Machine Learning Authors: Ran Zhou, Zehui Kong, Fanyu Lin, Jiaheng Liu, Zhen Zhang, Yuxiang Ren Published: 2026-10-22 arXiv: 2610.26995

Abstract

LLM-based agents increasingly rely on reusable skills. However, scaling skill libraries makes efficient and accurate skill use difficult. Full skill loading incurs high token and runtime overhead. In contrast, flat semantic retrieval treats skills as independent text entries, ignoring dependencies, execution order, state transitions, and data flow constraints.

We propose SPEM, a lightweight structure-aware procedural memory framework for efficient skill use in LLM-based agents. SPEM maintains Atomic Skill Memory for individual skills and Skill-DAG Memory for reusable compositions derived from successful executions.

For a new task, SPEM reuses high-confidence Skill-DAG templates when available. Otherwise, it retrieves atomic skills and invokes Skill-GNN for state- and structure-aware skill selection. Successful executions are written back as new Skill-DAGs, enabling procedural memory to accumulate over time.

Results

Experiments on ALFWorld and SkillsBench demonstrate that SPEM achieves the best overall effectiveness-efficiency trade-off:

ALFWorld: 87.1% success rate
SkillsBench: 38.38% reward
Reduced token and runtime costs compared to baselines