Google Research has introduced two complementary pieces of work aimed at giving AI models a form of long-term memory. The first, called Titans, is a new model architecture that moves beyond the standard transformer's fixed-size context window. Instead of treating every input as a fresh batch of tokens, Titans maintains a persistent memory that can be written to and read from over time.
The second piece, MIRAS, is a training method that teaches the model how to use that memory effectively. Rather than storing everything indiscriminately, MIRAS learns to decide which information is worth remembering and when to retrieve it. This makes the memory both efficient and relevant to the task at hand.
According to the researchers, combining Titans with MIRAS allows models to handle much longer sequences and to perform tasks that require recalling information from far back in the conversation or document. The work is still early, but it points toward a future where AI systems are not limited by a short-term buffer but can build up a durable understanding over time.