Associative Recurrent Memory Transformer
Summary
Associative Recurrent Memory Transformer extends RMT with layerwise associative memory for much longer segmented contexts.
Role In The Wiki
Use this page as the object card for ARMT. The source page carries the evidence details, limitations, and agenda mapping.
Relation To Foundation TSFM Agenda
ARMT is architecture background for long-context memory capacity and mutable key/value state. It is promising as a memory primitive, but the source page should carry the caveat that the reported evidence is language and associative retrieval rather than numeric time-series or action-conditioned control.
Evidence
Official Artifacts
- Official code: RodkinIvan/associative-recurrent-memory-transformer
Overlap Notes
ARMT is the RMT-family capacity branch. Its associative memory makes it closer to MIRAS and Titans than plain RMT is, but its update contract is still segment-level recurrence rather than a general test-time memory module.
For time-series work, ARMT should be compared against Mamba and RWKV-TS as different ways to retain long history under bounded serving cost. The missing evidence is still numeric time series, event streams, and action-conditioned dynamics.