[HUGGINGFACE]score: 0.55
AURA-Mem Adds Constant-VRAM Recurrent Memory to VLA Robot Policies
June 1, 2026
AURA-Mem wraps a frozen vision-language-action backbone with a fixed-size recurrent memory and a learned action-utility gate that only writes when the current observation would change the next action, keeping VRAM usage constant across arbitrarily long episodes. The design targets edge robotics hardware where KV-cache growth and flash write endurance are binding constraints.
paper
HOW THIS AFFECTS YOU
●
builderConstant-VRAM memory for long-horizon robot episodes is a practical constraint solver for edge deployment — the gating mechanism avoids unbounded cache growth without retraining the backbone.
●
researcherAction-conditioned gating as a memory write criterion is a novel design choice that decouples memory size from episode length in embodied agent architectures.