Titans
Summary
Titans tracks the neural-memory sequence-model family that pairs attention-like short-term context with learned long-term memory updated at test time.
Role In The Wiki
Use this page as the object card for the model or method. The source page carries the evidence details, limitations, and agenda mapping.
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in titans-2025 rather than duplicating verdict rows here. At the entity level, this page should stay as the object card; source pages carry slot-level evidence, limitations, and missing pieces.
Evidence
Overlap Notes
Titans is the learned test-time-memory branch. It differs from RMT because its long-term memory is a separate neural memory updated during inference, while RMT carries explicit memory tokens between sequence segments.
ATLAS is the direct higher-capacity successor in this wiki. MIRAS is the framework page for comparing Titans-style memory with attentional bias, retention, and online learning rules. MesaNet is adjacent because it spends inference compute on local optimization rather than only updating a memory module. Titans Revisited is the cautionary baseline page and should be checked before treating Titans-style memory as settled evidence.