Researchers have introduced SimpleMemVLA, a vision-language-action (VLA) model that tackles long-horizon robot manipulation without any dedicated memory module. The work, posted on Hugging Face, argues that the problem is fundamentally one of partial observability: the observation needed to pick the next action may have appeared only minutes earlier.
Existing approaches — retrieval banks, learned compressors, and recurrent states — must decide what to retain from the past before knowing what a future decision will require. That design, the authors note, rested on the assumption that minute-scale history is too large to process directly, an assumption they say modern VLM backbones no longer justify.
SimpleMemVLA instead keeps the sampled history intact and passes it to the backbone in the timestamped video format the backbone was pretrained to handle. The hidden states of a generated sub-task then serve as the only channel from history to a standard flow-matching action head.
To keep inference practical, the method exploits the fact that consecutive decisions share most of their history: prefilling the shared prefix during action execution holds latency close to that of a single-frame VLA.
The authors report that SimpleMemVLA sets a new state of the art on four memory benchmarks without sacrificing general-purpose control. With the backbone and training setup held fixed, it outperforms retrieval, compression, and recurrent-state mechanisms by a wide margin, and causal interventions are said to confirm that the policy genuinely reads its history. Code is available on GitHub.