What's this about?
Autoregressive video diffusion generates long video in chunks from a bounded recent context. Recency-based caching keeps local continuity, but it evicts the historical cues you need when a subject, object, scene or attribute *reappears*. LayerRecall is a state-conditioned memory router for long-horizon consistency. 26 upvotes.
Why is reappearance the hard case?
Because keeping the last few seconds is enough for smooth motion and useless for identity. A character walks out of frame at second five and returns at second forty; by then everything about them has been evicted, so the model regenerates them from a text description and they come back subtly wrong — different jacket, different face, different room behind them. Fixing it requires deciding what to retain by *what it is* rather than *when it happened*, which is a routing problem, not a capacity problem.
Why does this matter to you?
- It is the same argument as the Compaction Cliff for text agents: age is the wrong eviction criterion when some content must survive verbatim.
- It is the mechanism behind the 40-second scene-extension race that Gemini Omni and others are running.
- "State-conditioned router" is the general shape of an answer here — decide what to keep based on what is currently happening.