Recurrent Action Transformer with Memory
Summary
Recurrent Action Transformer with Memory is an offline RL policy architecture that adds recurrent memory embeddings and a Memory Retention Valve to Transformer trajectory modeling.
Role In The Wiki
Use this page as the object card for RATE. The source page carries the evidence details, limitations, and agenda mapping.
Relation To Foundation TSFM Agenda
RATE is the direct action-trajectory member of the RMT family in this ingest batch. It is useful for thinking about action histories, partial observability, and sparse delayed cues, but it should not be described as an explicit action-conditioned world model unless paired with a learned dynamics interface.
Evidence
Official Artifacts
- OpenReview: ICLR 2026 Poster
- Official code: CognitiveAISystems/RATE
- Project page: RATE model
Overlap Notes
RATE is the action-trajectory member of the RMT family. It overlaps with World Models because it has observations, actions, rewards, and memory under partial observability, but the wiki boundary is important: RATE chooses actions from offline trajectories rather than learning a transition model for candidate-action rollout.
Compared with ARMT, RATE is less about memory capacity over arbitrary key/value associations and more about retaining decision-relevant cues across trajectory chunks. Its Memory Retention Valve should be compared with test-time-memory retention mechanisms, but the claim remains policy-side evidence unless paired with a learned dynamics model.