Addressable Memory for Video World Models

TL;DR - WorldTrace is a training-free KV-cache memory framework that keeps interactive video world models able to "remember" and re-render previously seen scenes far beyond their training horizon. It matters because long-horizon visual persistence is a core blocker for interactive/generative world models, and this fixes it without retraining.

• Diagnoses the failure mode: once rollouts exceed the training horizon, temporal RoPE offsets go out-of-distribution, so attention can no longer address stored frames in the growing KV cache; naive compression in RoPE-rotated space further corrupts memory by averaging incompatible positional phases.
• Core fix: assign each compressed summary slot a distinct, in-distribution virtual position, keeping the compacted cache addressable.
• Two compression variants: WorldTrace-Field compresses history for temporal coherence; WorldTrace-Landmark stores verbatim scene traces at detected transitions for episodic recall.
• Introduces LoopBench (can a compressed cache reconstruct a previously visited scene after a long detour); reports +15.5% temporal consistency (Field) and +19.5% episodic recall (Landmark), all without retraining.