WorldTrace Fixes the Bug Behind Video Models' Forgetting

Video world models lose memory once rollouts run past their training horizon, per the WorldTrace paper. That's a problem for anyone building agents or simulators that depend on long, coherent rollouts. WorldTrace's fix recovers up to 19.5% of that lost episodic recall on the new LoopBench benchmark.

The root cause is RoPE, the positional encoding these models use to track time across frames. Long rollouts push those offsets past the range the model trained on, so attention can't reliably address older stored content.

Naive cache compression makes it worse, per the paper. It averages positional phases that don't match, scrambling the slots meant to hold memory.

WorldTrace's fix skips retraining. It gives each compressed memory slot its own valid position instead of an averaged one. Two versions test the idea: Field, tuned for temporal coherence, and Landmark, tuned for verbatim episodic recall. On LoopBench, Field improved temporal consistency 15.5% and Landmark improved episodic recall 19.5%.

The forgetting looks like a hard memory ceiling, but it's actually a bug in how compressed positions get indexed, fixable without retraining a single weight. If that holds up, other long-context video claims are worth re-checking against how their caches assign position, not just how large the advertised window is.