Generated illustration for VLAs with Long and Short-Term Memory — Paper 12 of 41Seedream 4.0 · GMI Studio

Illustration brief the model wrote

A robot arm in a kitchen placing a pot on a stove. To the left, a floating screen showing a rapid sequence of video frames from the last 30 seconds. To the right, a floating notepad with bullet-point text summaries of earlier steps. No text, no letters, no human faces.

Provenance. Brief written by deepseek-ai/DeepSeek-V4-Flash via GMI Cloud on 2026-09-18 from the paper's title, primary link and the editor's notes; reviewed editorially. Graphic rendered by seedream-4-0-250828 through the GMI Studio request queue (request 99388ffc…), converted to PNG locally. Summaries can be wrong; the linked paper is the source of truth.

Atlas/Physical Intelligence·Physical AI company·2026

VLAs with Long and Short-Term Memory

Model tl;dr

Multi-scale memory stores video clips and language summaries, enabling robots to complete 15-minute tasks and recover from errors.

embodied memorymulti-scalevideo language actionlong-horizonerror recoveryrobot manipulation

What changed

Model

Before this paper, robot policies with VLAs were limited to short tasks because they lacked persistent memory across time. This paper introduces a memory system that stores short-horizon video frames and long-horizon language summaries, allowing tasks up to 15 minutes long and in-context recovery from earlier mistakes.

So what, in plain English

Model

Think of a robot with a notepad where it quickly sketches what it sees every few seconds and writes a longer diary entry every few minutes, so it can remember the whole cooking session. This means a robot can prepare a full meal without needing to start over if it drops a spoon.

Contribution to industry

Model

This work introduces a practical memory architecture for VLA-based robot policies, enabling long-duration tasks (15 minutes) in commercial cleaning, service, and kitchen settings. It reduces the need for human resets and supervision, moving robots closer to viable deployment in messy, real-world environments.

Typed judgments

Jev

★★★☆☆ robots · Physical AI 95%

Real hardware 51% · Open code / data 8% · Foundation-model based 86% · Survey / benchmark-only 6%

Probabilities from TypeSafe Jev over the title, venue and abstract; the filters on the atlas apply code-owned thresholds to these numbers.

Editor's notes

Summary. Multi-scale Embodied Memory keeps short-horizon video history and long-horizon language summaries that the model chooses to store, enabling tasks up to 15 minutes long and in-context recovery from earlier mistakes.

Industry. Lifts the short-task ceiling on robot policies, a prerequisite for commercial cleaning, service and kitchen jobs that run for many minutes without resets.

Read the paper ↗Official research index ↗All 8 Physical Intelligence papers →