Summary
Action-conditioned world models that generate multi-segment videos frequently fail at memory: after the camera leaves a region and returns, scenes or salient objects silently change. Existing memory designs are hard to compare because gains are entangled with backbone, training, retrieval, and evaluation differences. Echo-Memory fixes the action-to-video interface and varies only how history is stored and read, providing the first controlled ablation of memory mechanisms in world action models.
Key Contributions
- Controlled experimental framework isolating memory design from backbone and training confounders
- Systematic ablation of how history is stored (episodic buffer, compressed tokens, etc.) and read (attention, retrieval, gating)
- Identifies memory as the primary failure mode for scene consistency in multi-segment video world models
- Provides design guidelines for future world action model memory architectures
Significance
A rare controlled study in a field where ablations are often confounded by architectural choices; the findings provide principled guidance for memory design in robot world models.