Chen-Led Survey Maps Memory Mechanisms in AR Video Generation

Autoregressive video generation, the paper writes, 'extends visual sequences through causal rollouts.' Then the models run out of room. 'Critical historical information, e.g., entity identities, dynamic states, and intervention-induced causal changes, often leaves the active context long before its relevance diminishes.'

That is the gap a new Hugging Face-listed survey, 'The Past Frames the Future: Memory for Autoregressive Video Generation,' sets out to organise. Lead author Harold Haodong Chen and his co-authors reframe the shortfall as 'a fundamental memory problem' and pull the scattered literature under one roof.

The taxonomy runs across five perspectives: 'Forms, the representational carriers of history; Functions, the specific semantic and physical information requiring preservation; Operations, the lifecycle of writing, reading, updating, managing, and integrating memory; Learning, the optimization of memory behaviors under closed-loop rollouts; and Evaluation, the paradigms for diagnosing genuine memory capabilities.' Memory itself is defined operationally as 'persistent historical information maintained across outer AR steps, capable of influencing future generation even after the originating evidence is no longer locally accessible.'

The closing note flags the open gaps: 'composable and resource-aware memory architectures, trustworthy state updating, self-rollout learning, and standardized evaluation.' No benchmarks are proposed, no headline numbers are offered. It is a mapping exercise, one that lands amid a steady run of research alerts on our tracker where memory has been a repeat character, from Jev-Mem's LoCoMo latency cut at UT Dallas to the Kimi Delta attention work.