How procedural memory can cut the cost and complexity of AI agents

Researchers from Zhejiang University and Alibaba Group developed Memp, a framework that gives AI agents "procedural memory" similar to human skill learning. This system allows agents to continuously build, retrieve, and update memories from past experiences, making them more efficient at complex tasks. Unlike other memory frameworks that focus on remembering conversation content, Memp captures reusable "how-to" knowledge across different tasks. The framework addresses the cold-start problem by using evaluation metrics to bootstrap initial memories and demonstrates transferability between large and small models, potentially reducing costs while maintaining performance.