2AM Decouples Task Memory from Action Models in Robotics
September 11, 2026
2AM uses a multimodal Agent to hold all task memory, passing subtask language and 2D spatial hints to a stateless, RGB-based Action Model. This design isolates the source of errors between the high-level language interface and the low-level motor execution in long-horizon manipulation.
HOW THIS AFFECTS YOU
●
builderYou can implement steerable robotics by moving memory management from the policy to a separate agentic layer.
●
researcherThis architecture allows for better attribution of failures in complex manipulation tasks.