Rethinking Memory Mechanisms of Foundation Agents in the Second Half

A major survey by 59 authors argues the AI field's next challenge is creating agents that work in complex, real-world scenarios. This requires sophisticated memory systems to manage vast information over time. The paper organizes memory research into three dimensions: physical form, cognitive type, and focus. It analyzes how memory is built and used in different agent designs, reviews evaluation methods, and outlines critical open problems for future development.