A new memory layer lets existing robot control models remember what they did a few steps ago, without retraining them.
Researchers built SimpleARM, a training-free memory module that sits on top of frozen generalist robot policies. It reads the task instruction to decide what to track, uses off-the-shelf perception tools to keep a compact log of relevant state (which object is which, how far along a multi-step task is, what order steps happened in), and pulls that log back in only when the robot's next move actually depends on history. Tested on RoboMME, a 16-task benchmark built specifically to require memory that a single camera frame cannot supply, SimpleARM produced a 67.17% mean success rate across three policy seeds. The strongest non-oracle baseline reached 44.51%.
That is a real gain, roughly 1.5x, not the near-doubling some coverage of this paper implies, and it arrives without touching the underlying policy's weights. Most robot memory approaches either retrain the policy or hoard raw video frames, both of which get expensive fast. SimpleARM's ablations show each type of state (identity, progress, sequence) only matters on the tasks that actually need it, a cleaner result than this kind of paper usually delivers.
Robotics is relearning a lesson language models absorbed a few years back: retrieval that knows what to fetch beats memory that just piles up everything it has seen.