SimpleMemVLA: A Simple but Effective Native-Video Memory for Vision-Language-Action Models

Hugging Face Daily Papers Papers

Summary

SimpleMemVLA introduces a simple memory mechanism for Vision-Language-Action models by feeding intact timestamped video history into a pretrained VLM backbone, achieving state-of-the-art results on long-horizon manipulation tasks without dedicated memory modules.

Long-horizon manipulation is partially observable: the information needed to choose the next action may appear only in observations from minutes earlier. Existing memory mechanisms: retrieval banks, learned compressors, recurrent states must decide what to keep from the past before knowing what a future decision will require. This was motivated by the assumption that minute-scale history is too large to process directly, which modern VLM backbones no longer make true. In this work, we introduce SimpleMemVLA, a VLA without a dedicated memory module. It keeps the sampled history intact and passes it to the backbone in the timestamped video format the backbone was pretrained to process; the hidden states of a generated sub-task then form the only channel from history to a standard flow-matching action head. Since consecutive decisions share most of their history, prefilling the shared prefix during action execution keeps latency close to a single-frame VLA. SimpleMemVLA sets a new state of the art on four memory benchmarks without cost on general-purpose control. Holding the backbone and training setup fixed, it outperforms retrieval, compression and recurrent-state mechanisms by a wide margin, and causal interventions confirm that the policy genuinely reads its history. Code available at https://github.com/wadeKeith/SimpleMemVLA
Original Article
View Cached Full Text

Cached at: 09/09/26, 04:31 AM

Paper page - SimpleMemVLA: A Simple but Effective Native-Video Memory for Vision-Language-Action Models

Source: https://huggingface.co/papers/2609.05533 Authors:

,

,

,

,

,

,

,

,

,

Abstract

SimpleMemVLA achieves long-horizon manipulation by feeding intact timestamped video history directly into a pretrained VLM backbone and using hidden states to inform a flow-matching action head, outperforming dedicated memory modules.

Long-horizon manipulation is partially observable: the information needed to choose the next action may appear only in observations from minutes earlier. Existing memory mechanisms: retrieval banks, learned compressors, recurrent states must decide what to keep from the past before knowing what a future decision will require. This was motivated by the assumption that minute-scale history is too large to process directly, which modern VLM backbones no longer make true. In this work, we introduce SimpleMemVLA, aVLAwithout a dedicated memory module. It keeps the sampled history intact and passes it to the backbone in thetimestamped videoformat the backbone was pretrained to process; thehidden statesof a generated sub-task then form the only channel from history to a standardflow-matching action head. Since consecutive decisions share most of their history, prefilling the shared prefix during action execution keeps latency close to a single-frameVLA. SimpleMemVLAsets a new state of the art on fourmemory benchmarkswithout cost on general-purpose control. Holding the backbone and training setup fixed, it outperforms retrieval, compression and recurrent-state mechanisms by a wide margin, andcausal interventionsconfirm that the policy genuinely reads its history. Code available at https://github.com/wadeKeith/SimpleMemVLA

View arXiv pageView PDFProject pageGitHub1Add to collection

Get this paper in your agent:

hf papers read 2609\.05533

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.05533 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.05533 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.05533 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles