Long-form agentic video generation struggles to keep a screenplay visually coherent when shots are requested out of order. Simply passing prompts or recent frames between shots gives the model no shared spatial frame for revisiting a scene and no time-indexed record for replaying an earlier event. The proposed ATAS framework addresses this by treating generation as a world-state problem rather than a frame-to-frame problem.

ATAS first compiles a screenplay into shot requests that carry explicit location, camera, and event-time references. Before generating a clip, it retrieves relevant evidence from two memories: a scene memory of registered static geometry and an event memory of dynamic observations indexed by scene, event, and time. After generation, the clip is reconstructed and registered, but memory is updated only when review and state checks accept the observation. Film delivery remains independent of memory admission, so rejected observations do not block output.

In experiments across four video generators, ATAS improved all reported cross-shot consistency measures and achieved the highest aggregate benchmark score for each backbone. It also outperformed compared agent-based and 3D-guided methods on target-image spatial revisit, led on three event-recapture criteria, and earned the highest human ratings on all six evaluation criteria. The authors have released code and video results on a project page.