ME AI message: Physical Intelligence and multiple universities have proposed a Multi-Scale Embodied Memory (MEM) method to address the memory bottleneck in robots’ long-horizon tasks. The method uses a dual-modal memory mechanism: video frames are used to store short-term memory, natural language is used to abstract and store long-term concepts, and a reasoning mechanism proactively decides what to remember and how to remember—enabling coordinated inference between “what to do” and “what to remember.” Experiments show that MEM allows robots to complete multi-stage tasks lasting up to 15 minutes, such as cleaning the kitchen or making a grilled cheese sandwich, and to correct mistakes during execution. The study indicates that poorly designed memory systems can cause “causal confusion,” which intensifies false correlations in imitation learning; by proactively maintaining the text memory stream, MEM can track an object’s position and state even when it is out of view. The paper has been publicly released. (Source: ME)