Agent-side memory steers stateless robots through long manipulation tasks

2AM separates memory from motor execution in long-horizon manipulation. A multimodal agent keeps all task history and, each step, steers a stateless RGB-based action model with subtask language plus optional 2D grasp, place, and move hints, rather than using depth, online geometry, or a planner. The action model was trained on demonstrations augmented with hint labels, then regularized with hint dropout, spatial noise, and temporal jitter so it can follow imperfect agent outputs. On LIBERO-Mem, 2AM reached 76.3% average completion against a 14.8% baseline, with 63.0% relaxed and 11.8% strict success. The result is interesting because it isolates the source of long-horizon gains: the agent's memory and steering matter independently of policy memory, richer observations, or geometric tools. It also suggests an action model's capability is gated by how precisely it is steered, not only by what it learned.