What Is MemoryVLA?

MemoryVLA is introduced in arXiv 2508.19236 by Hao Shi, Bin Xie, Yingfei Liu, Lin Sun, Fengrong Liu, Tiancai Wang, Erjin Zhou, Haoqiang Fan, Xiangyu Zhang and Gao Huang (v1 26 Aug 2025, v2 30 Jan 2026). The project page lists Tsinghua University, Dexmal, MEGVII, TJU, HIT and StepFun. Read: the arXiv abstract page, project page and repository README; the full paper text was not read.

Third-party research; Firsthand is not affiliated and claims no use, training or compatibility with it. Firsthand captures first-person human video, not robot data, fast, on demand and to a written spec: 21 keypoints per hand with a visibility flag, wrist trajectories and action segments, plus one task instruction per episode.

The abstract says temporal context is essential for robotic manipulation because such tasks are inherently non-Markovian, yet mainstream VLA models typically overlook it and struggle with long-horizon, temporally dependent tasks. The project page gives a Push Buttons example, where the states before and after a push look nearly identical.

The abstract describes a Cognition-Memory-Action framework. A pretrained VLM encodes the observation into perceptual and cognitive tokens that form working memory. A Perceptual-Cognitive Memory Bank stores low-level details and high-level semantics. Working memory retrieves decision-relevant entries, fuses them with current tokens and updates the bank by merging redundancies. A memory-conditioned diffusion action expert then yields temporally aware action sequences.

The project page adds a 7B VLM, cross-attention retrieval with timestep positional encoding, a gate for fusion, and a merge of the most similar adjacent entries when the bank reaches capacity. The capacity value is Not stated here.

The project page says the model takes an RGB observation and a language instruction. The README lists MemoryVLA on the OpenVLA codebase, MemoryVLA+ on the Dexbotic codebase and MemoryVLA++, the extended journal version, whose code release is listed as pending. Training uses RLDS-format LIBERO, Bridge and Fractal data, starting from an OpenVLA checkpoint for LIBERO and a CogACT checkpoint for Bridge and Fractal, on a single node with eight A100 GPUs. For custom data the README asks for RLDS with end-effector delta actions plus gripper open or close.

The abstract reports evaluation on 150+ simulation and real-world tasks across three robots. It gives success rates of 71.9% on SimplerEnv-Bridge, 72.7% on Fractal, 96.5% on LIBERO-5 and 41.2% on Mikasa-Robo, which the authors say outperform CogACT and pi-0, and 84.0% on 12 real-world tasks, with a +26 improvement on long-horizon tasks. These are the authors’ numbers, and this page repeats no baseline scores.

The project page groups real-world tests into a Long-horizon Temporal suite and a General suite. The long-horizon list includes Sequential Push Buttons, Change Food, Guess Where, Clean Table & Count, Pick Place Order and Clean Restaurant Table. It also shows robustness variations for two tasks, such as unseen background, lighting and occlusion. Per-task scores are Not stated here.

The site’s entries describe CronusVLA as adding multi-frame history through feature chunking, and a long-horizon task as a multi-step job captured with ordered action segments. MemoryVLA is a model design, while a long-horizon task is a property of the data. The table sets them side by side and ranks none of them.

The full paper’s method detail, the Hugging Face model cards and the MemoryVLA++ paper were not read. The venue differs between sources: arXiv and the README say ICLR 2026, while the project page header says CVPR Workshop Oral. The README text read states no licence, so terms for the code and datasets are Not stated here. MemoryVLA’s stated data is robot trajectories with images and instructions, and the sources do not say whether first-person human video would suit it.

MemoryVLA’s stated inputs are robot data, so human video is not its stated input. For long-horizon work, Firsthand records each full attempt as one episode with ordered, timestamped action segments, plus hand pose, wrist trajectories and one task instruction per episode. Robot actions are Not stated. Write tasks, environments and signals into a spec and confirm on a Spec Pilot. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Formats include RLDS; matching MemoryVLA’s expected fields is the buyer’s to check.

What is MemoryVLA?

A vision-language-action framework (arXiv 2508.19236, ICLR 2026) that adds a Perceptual-Cognitive Memory Bank so a model can use temporal context on long-horizon manipulation.

What is the Perceptual-Cognitive Memory Bank?

Per the abstract, a store of low-level details and high-level semantics that working memory queries, fuses with current tokens and updates by merging redundancies.

How is MemoryVLA different from CronusVLA?

Per the abstracts, MemoryVLA retrieves from a memory bank, while CronusVLA aggregates history through feature chunking. This page ranks neither.

Is MemoryVLA open source?

The official repository releases code and links model weights and datasets on Hugging Face. The README text read states no licence, and the model cards were not opened here.

Does Firsthand use MemoryVLA, and can I request a custom dataset?

Firsthand is not affiliated and claims no use, training or compatibility. You can request a custom dataset, confirmed first on a Spec Pilot.