The paper introduces Relax Forcing, a training‑free memory mechanism for autoregressive video diffusion that structures temporal context into Sink, Tail, and History frames. By selecting History frames with a relaxation criterion, the method reduces error accumulation and attention overhead while preserving motion dynamics. Experiments on VBench‑Long demonstrate that this structured memory improves long‑video generation quality over existing baselines.
By Zengqun Zhao, Yanzuo Lu, Ziquan Liu, Jifei Song, Jiankang Deng, Ioannis Patras
arXiv:2606. 09803v1 Announce Type: cross Abstract: We present \textbf{Echo-Memory}, a controlled study of memory mechanisms in action-conditioned world models.
By Wayne King, Zeyue Xue, Yuxuan Bian, Jie Huang, Haoran Li, Yaowei Li, Yaofeng Su, Yuming Li, Haoyu Wang, Shiyi Zhang, Songchun Zhang, Yuwei Niu, Sihan Xu, Junhao Zhuang, Haoyang Huang, Nan Duan
arXiv:2610.35001v1 Announce Type: new
Abstract: Long-video generators increasingly retrieve past visual observations to prevent identity drift. The intuition is compelling: more access to history sho...
By Yuzhuo Chen, Huafeng Shi, Xinyu Wang, Yulong Xu, Yucheng Wang, Haoqin Hong, Guoxin Zhang, Zehua Ma, Weiming Zhang
LayerRecall is a memory router for autoregressive video diffusion that selectively retrieves and injects historical key/value states into specific layers of the model, based on the current context. It addresses the problem that existing memory mechanisms expose nonlocal history but do not guarantee effective use, by recognizing that different layers prefer current, recent, or distant context. The method, combined with Cross‑Horizon Prediction Matching, achieves state‑of‑the‑art long‑range consistency on MemoBench and MovieBench while maintaining local continuity and incurring negligible inference overhead.
By Yixuan Ding, Jiahao Kong, Wei Huang, Ruijie Quan, Yi Yang
Long-video generators increasingly retrieve past visual observations to prevent identity drift. The intuition is compelling: more access to history should make future video more consistent. But an ima...
MemInStream is a new benchmark designed to evaluate memory capabilities in streaming video understanding. It introduces an evidence‑controlled validation protocol that links each question to temporally localized evidence intervals, ensuring that models must rely on historical memory rather than recent observations or linguistic priors. The benchmark covers four key categories—Episodic Retrieval, State Tracking & Updating, Multi‑Evidence Integration, and Contextual Memory—across diverse video scenarios to systematically assess memory performance.
By Zhenqi He, Dunyuan Xu, Yang Bo, Kien T. Pham, Yanghao Wang, Shuhan Li, Xintong Guo, Yuyang He, Fang Liu, Huankang Guan, Jingyan Chen, Rui Liu, Long Chen
The paper introduces Prediction‑Aligned Context Compaction (PACC), a method that learns a compact memory representation for long‑video generation by distilling a frozen video generator. PACC trains a compressor to aggregate past frames into memory tokens, using the generator as both teacher and student during on‑policy distillation. Experiments on MBench and VBench‑Long show that PACC improves memory‑event coverage and consistency, achieving better scores than strong baselines and producing competitive minute‑long videos.
By Xiaoyu Wu, Weihang Guo, Yifei Wang, Xinze Feng, Lydia E. Kavraki, Zhiwei Steven Wu
EGOSTREAM is a diagnostic benchmark for evaluating streaming episodic memory in egocentric vision, comprising 2,250 curated questions across seven cognitive dimensions and expanding to 8,528 recall‑conditioned evaluations using the Answer Validity Window. It enables controlled testing from instant to ultra‑long‑term recall while distinguishing genuine forgetting from natural scene changes. Baseline experiments on a unified MLLM framework reveal that different memory‑management mechanisms yield similar overall accuracies but markedly different memory profiles, with all in‑cache policies operating below real‑time and achieving at most 45% accuracy under a 9,804‑token budget.
By Rosario Forte, Giuseppe Lando, Antonino Furnari
The paper introduces Learning-to-Retrieve (L2R), a method that embeds memory retrieval directly into video world models. By using the model’s internal state as a memory and employing a camera‑conditioned retrieval gate and a retrieval trigger supervised by a 3D re‑visibility signal, the approach learns when and what historical information to access. This eliminates the need for external memory systems and improves long‑term scene consistency across various base models and camera‑revisit benchmarks.
By JiaKui Hu, Tailai Chen, Yuqi Pan, Xuerui Qiu, Jialun Liu, Xiao Cao, Zhenxin Zhu, Guang Chen, Hangjun Ye, Bing Wang, Yanye Lu
arXiv:2609.37559v1 Announce Type: new
Abstract: To serve as real-world personal assistants, streaming video models need persistent memory that retains past experiences for later use. Yet existing str...
By Jianguo Huang, Jinming Liu, Qiyao Wang, Liang Xu, Jianhang Li, Zhimian Wen, Mingda Li, Shule Lu, Zhicheng Wang, Yuhan Guo, Xin Jin, Wenjun Zeng
GROME introduces a factorized memory framework for video world models that separates historical evidence into static and dynamic components. It uses local memory retrieval to align static history spatially and injects the factorized memory into models without extra training. Experiments show improved long‑horizon scene persistence and dynamic entity fidelity while maintaining visual quality and controllability.
By Fang Yu, Xiaoxue Wu, Haiyu Zhang, Yangfu Li, Botong Zhao, Yu Qiao, Tianfan Xue, Yue Lu, Xinyuan Chen
arXiv:2605.25333v3 Announce Type: replace
Abstract: Video world models should maintain evolving states when evidence is unobserved, yet current generators often freeze hidden states upon interruption...
By Tianshuo Xu, Yichen Xie, Depu Meng, Chensheng Peng, Quentin Herau, Bo Jiang, Yihan Hu, Wei Zhan