$\mu$VLA: On Recurrent Memory for Partially Observable Manipulation in VLA Models

Vision-language-action (VLA) models predict chunks of future actions from the current observation, an assumption that fails under partial observability, where decisions depend on information no longer visible. Existing memory-augmented VLAs simultaneously introduce recurrence, retrieval, compression modules, auxiliary objectives, hierarchical memory, or task-specific architectural changes, so the contribution of recurrence itself remains entangled with surrounding machinery. We present a controlled isolation study of recurrence in a strong pretrained VLA backbone. Our formulation augments the transformer with a small set of learnable memory tokens carried across timesteps and updated through self-attention, trained end to end with truncated backpropagation through time, with no auxiliary losses and no architectural changes. We instantiate this as $\mu$VLA, a family of OpenVLA-OFT variants parameterized by memory width $m$, TBPTT length $K$, and the memory update rule (cross-step gradients or a detached EMA), so that recurrence is the only varying factor. On MIKASA-Robo-VLA, $\mu$VLA improves average success rate on five training tasks from $0.42$ to $0.84$ at the strongest setting and reaches $0.23$ on held-out tasks with the same memory structure versus $0.07$ for the memoryless baseline, while staying near baseline on tasks requiring different memory structure. On LIBERO, the strongest recurrent variant achieves $95.9\%$ average success, indicating no regression under full observability. These results calibrate the capability envelope of minimal in-backbone recurrence: the regime where it suffices, and the regime requiring additional memory structure. Under open-loop chunking the same write reaches mean SR $0.30$ on eight further tasks. Demos: https://anon123p.github.io/anonymous-site/.