SimpleMemVLA: A Simple but Effective Native-Video Memory for Vision-Language-Action Models
The paper argues that long-horizon robot control can use raw sampled video history directly, without a separate memory module.
SimpleMemVLA feeds timestamped past observations to the pretrained vision-language backbone in its native video format. The generated sub-task hidden states are then the only path from that history into a standard flow-matching action head. The authors say prefix prefill keeps latency close to a single-frame VLA because consecutive decisions reuse most of the same history. They report new state-of-the-art results on four memory benchmarks, with no loss on general-purpose control. HF Daily Papers' note
SimpleMemVLA feeds timestamped past observations to the pretrained vision-language backbone in its native video format. The generated sub-task hidden states are then the only path from that history into a standard flow-matching action head. The authors say prefix prefill keeps latency close to a single-frame VLA because consecutive decisions reuse most of the same history. They report new state-of-the-art results on four memory benchmarks, with no loss on general-purpose control. HF Daily Papers' note
score 5