NativeMEM: Native Memory Compression for Long-Horizon Robotic Manipulation

2026-07-07

AI Summary

This work addresses the challenge that existing pretrained vision–language–action (VLA) models struggle to efficiently retain long-horizon visual history during high-frequency inference. The authors propose NativeMEM, a novel approach that repurposes the VLA model’s own visual encoder to compress multi-view historical frames into a single memory token, which is then integrated directly into the input sequence. This enables native access to long-term memory without requiring external memory modules or planners. NativeMEM employs a two-stage training strategy: first learning a general-purpose memory tokenizer under supervision from a frozen VLA, followed by task-specific fine-tuning. Experiments demonstrate that NativeMEM boosts success rates in simulation from 32.4% to 84.0% and achieves 98.7% on real robots, outperforming prior methods with only 20% of the training data while maintaining low inference latency and GPU memory consumption.

Abstract

How can pretrained Vision-Language-Action (VLA) models retain long-horizon visual histories with high-frequency updates without sacrificing efficiency? Existing approaches rely on external memory management, which restrains either the memory horizon or the reactiveness of pretrained policies. To this end, we present NativeMEM, a VLA policy that features long-term and real-time updated memory. At its core is an efficient memory encoding scheme, Native Memory Compression, which repurposes the VLA's own vision encoder to compress each historical frame from each camera view into a single token. Appended to the input sequence, these memory tokens enable the pretrained VLA to attend over long-term history with negligible latency overhead, requiring neither an external planner nor a freshly initialized memory module. To align the memory tokens with the pretrained policy, we first develop a generic memory tokenizer under the supervision of a frozen VLA on memory-demanding data, and then unfreeze the VLA for task-specific fine-tuning. 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 maintaining low inference latency and GPU memory usage. Notably, NativeMEM exhibits high data efficiency by achieving competitive results with prior arts using only 20% of the training data.

Problem

• long-horizon manipulation
• visual history retention
• memory efficiency
• Vision-Language-Action models
• real-time memory

Innovation

• Native Memory Compression
• Vision-Language-Action (VLA)
• long-horizon manipulation
• memory-efficient inference
• pretrained policy alignment