Managing Procedural Memory in LLM Agents: Control, Adaptation, and Evaluation

The paper introduces AFTER, a benchmark comprising 382 enterprise tasks across six professional roles and 22 procedural skills, aimed at evaluating the transferability of skills in LLM agents. Results indicate that procedural memory enhances performance in industrial workflows, with a single refinement round yielding a 3.7-6.7 point improvement and achieving 73.1% cross-model test accuracy through multi-model execution traces. This research offers critical insights for practitioners on effectively implementing and assessing procedural memory systems in AI agent applications.