#deep-learning

Google Research introduced Titans, a new architecture that gives transformers a learnable long-term memory. Instead of forgetting everything outside the context window, Titans maintain a neural memory that updates itself at test time using gradient descent. Basically, it learns to memorize on the fly.