Memory-Augmented Video Generation

Memory-augmented video generation studies how video models can retain useful information beyond their immediate context, supporting consistent characters, scenes, and events over long clips or repeated scene visits. Approaches include selecting and compressing frames or attention-cache entries, learning compact memory tokens or entity slots, and using camera geometry or 3D structure to retrieve relevant spatial evidence. These methods are used in multi-shot storytelling and interactive world models, and are tested on long-video, camera-control, and scene-revisit benchmarks such as VBench-Long and revisit-consistency suites. A central challenge is keeping memory bounded without losing evidence, while allowing scenes and motion to evolve; replay fidelity alone may not measure whether a model can reliably recall a world.