Online reports that Google Research released a new study by Titans. By introducing a new neural long-term memory module, three-head collaborative architecture and hardware optimization design modules, the context window of the large model is expanded to 2 million tokens while increasing computing power by only 1.8 times. Titans not only solves the computing power bottleneck of the Transformer model in long-context processing, but also uses bionics to design a hierarchical mechanism that simulates the human memory system, realizing accurate reasoning of 2 million tokens of ultra-long contexts for the first time.