Pretrained Vision-Language-Action (VLA) models often struggle to keep long-horizon visual histories when updates arrive at high frequency. Existing approaches rely on external memory management, which tends to sacrifice either the memory horizon or the policy's reactiveness. The authors of NativeMEM propose a way to avoid that trade-off by repurposing the VLA's own vision encoder to compress each historical frame from each camera view into a single token.

These memory tokens are appended to the input sequence, allowing the pretrained VLA to attend over long-term history with negligible latency overhead. No external planner or freshly initialized memory module is needed. Training proceeds in two stages: first a generic memory tokenizer is learned under the supervision of a frozen VLA on memory-demanding data, then the VLA is unfrozen for task-specific fine-tuning.

The paper reports that NativeMEM consistently outperforms prior methods, boosting success rates from 32.4% to 84.0% in simulation and up to 98.7% on real robots, while keeping inference latency and GPU memory usage low. It also shows high data efficiency, achieving competitive results with prior arts using only 20% of the training data. The work was accepted to CoRL 2026; this summary is based solely on the arXiv preprint.