NativeMEM boosts robot success rates to 98.7% with native memory compression

NativeMEM, developed by a team including Ziye Wang and colleagues, tackles a core challenge in robotic manipulation: how to let pretrained Vision-Language-Action (VLA) models remember long visual histories without slowing down. Existing approaches often rely on external memory managers that limit either the memory horizon or the policy's reactiveness. NativeMEM's key innovation is a “native memory compression” scheme that uses the VLA's own vision encoder to compress each historical frame from every camera view into a single token. These tokens are appended to the input sequence, allowing the pretrained model to attend over long-term history with negligible latency overhead—no external planner or freshly initialized memory module required.

To align these memory tokens with the pretrained policy, the authors first develop a generic memory tokenizer under supervision of a frozen VLA on memory-demanding data, then unfreeze the VLA for task-specific fine-tuning. The results are striking: NativeMEM raises success rates from 32.4% to 84.0% in simulation and up to 98.7% on real robots, while maintaining low inference latency and GPU memory usage. Notably, it achieves competitive results with prior state-of-the-art using only 20% of the training data, highlighting its data efficiency. This work, published on arXiv (2607.06678), represents a practical step toward robots that can handle long-horizon tasks without expensive memory infrastructure.