PRISM
Scaling Short-Term Memory of Visuomotor Policies for Long-Horizon Tasks
Rutav Shah, Rajat Kumar Jenamani, Xiaohan Zhang, Lingfeng Sun, Roberto Martín-Martín, Yuke Zhu, et al. · 2026-06-15
- Management
- T2 Dynamic Memory
- Memory changes during deployment through a stable management policy.
- Representation
- Internalized latent & parametric
How experience becomes memory
Each path is one memory component, traced from what is stored to how it is found again.
hierarchically compressed visuomotor history
- Represented asPersistent multimodal latent stateInternalized latent & parametric
- Stored inHierarchical multi-tier store
- Retrieved byLatent / parametric activation
- Source evidenceSource not retained
Gated attention filters hierarchically compressed history tokens for action prediction.Full paper: Short-Term Memory · PDF pp. 4–8
How memory changes (T2 operations)
Update · Learned update module
During operation, PRISM applies a trained but frozen update module to hierarchically compressed visuomotor history; the memory state changes while the governing update policy remains fixed.