Yannic Kilcher Reviews Titans: Learning to Memorize at Test Time — Learnable Memory Modules Bridge Recurrent Models and Attention

Yannic Kilcher (YouTube)·medium signal

Kilcher's analysis of Titans (arXiv 2501.00663) covers a new architecture adding writable memory modules to transformers that enable memorization at test time, giving models persistent state they can update during inference without requiring larger context windows. The approach directly addresses long-horizon agentic task memory — the same problem class that MemPO and structured distillation papers have recently attacked from different angles.