This survey reviews how models maintain information beyond their bounded context window. It defines memory as persistent historical information that still influences future generation after the original evidence is no longer locally accessible. The authors organize existing techniques into five complementary perspectives: representational forms, semantic functions, operational lifecycle, learning methods, and evaluation paradigms. They argue that effective memory requires retained states to be accurate, accessible, and causally influential, not merely stored.
Persistent memory is essential for coherent autoregressive video generation
The authors organize existing memory mechanisms by representation, purpose, operation, training, and evaluation, while identifying requirements for reliable long-term generation.
Big Tech
Harold Haodong Chen · Rongjin Guo · Disen Lan · Wen-Jie Shu · Hongfei Zhang · Hanzhe Hu · +19 more
HKUST · CityUHK · FDU · ZODA · CMU
Research Digest··2 min read
In one line
Memory in autoregressive video generation is a five-part problem: forms, functions, operations, learning, and evaluation.
Before you read this
Three ideas this paper builds onautoregressive video generationGenerates video frames one at a time, each conditioned on previously generated frames.6 papers →memory retentionThe ability of a system to preserve and recall information over time or across contexts, often balancing compression against fidelity.32 papers →latent world modelA learned internal, compressed representation of environment dynamics, enabling future-aware planning, generation, and reasoning directly in latent space.51 papers →
Read it as
Chen and colleagues survey how autoregressive video generators can preserve relevant history after it leaves their limited context windows.
The authors review memory mechanisms for models that generate videos sequentially as tokens, frames, or chunks.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.