The “Surprise Metric”: Google’s New AI Architecture Outperforms GPT-4 in Memory

On 10 December 2025, Google unveiled two innovations—the Titans architecture and the MIRAS framework—designed to enable AI models to “read while remembering,” dynamically updating core memory in real time as they process vast information. The breakthrough aims to solve bottlenecks in traditional Transformer models when handling ultra-long sequences and demonstrably delivers strong performance in extreme long-context reasoning, managing contexts of up to two million tokens with remarkable ease.

A new long-term memory module, inspired by the brain’s separation of short- and long-term memory, underpins Titans, with a “surprise metric” guiding what information to store and what to forget. MIRAS provides a unified perspective and supports three model variants—YAAD, MONETA, and MEMORA—optimised for demands such as noise robustness and stable long-term memory retention.

In empirical evaluations, Titans and MIRAS-derived models outperformed leading architectures like Mamba-2 and Transformer++ in language modelling and commonsense reasoning, with the BABILong benchmark showing a dominance even though Titans uses far fewer parameters than GPT-4.