Google Research has introduced Titans, a new architecture designed to give AI models the ability to update their internal memory in real time, marking one of the most significant departures from the Transformer framework since its introduction in 2017.
The system, paired with a theoretical framework called MIRAS, is built to process and retain extremely long contexts while learning continuously during inference.
What Happened
The announcement addresses a longstanding limitation in large language models: while Transformers excel at pattern recognition, their computational cost scales poorly with longer inputs, and they cannot update core knowledge without retraining.
Google’s new approach allows models to modify their long-term memory parameters as data streams in, enabling persistent learning without offline fine-tuning.
According to Google Research, Titans combines the speed of recurrent architectures with the accuracy of attention-based systems, supported by a deep neural memory module that summarizes and integrates information across millions of tokens.
Also Read: PwC: Hedge-Fund Crypto Exposure Surges To 55%, Up From 47% Last Year
A key mechanism, described as a “surprise metric,” determines which new inputs meaningfully differ from the model’s existing memory and should therefore be stored permanently.
MIRAS, the accompanying blueprint, reframes all sequence models as variants of associative memory systems, defining how they store, retain, and update information.
The framework introduces several attention-free variants, including YAAD, MONETA, and MEMORA, each built to improve robustness or stability under long-context workloads.
Why It Matters
In experimental evaluations, Titans outperformed leading architectures such as Mamba-2, Gated DeltaNet, and Transformer++ on language modeling, zero-shot reasoning, genomics, and time-series tasks.
Google reports that Titans also achieved superior performance on the BABILong long-context benchmark, surpassing even GPT-4 despite having far fewer parameters, while scaling to context windows exceeding two million tokens.
Google positions Titans and MIRAS as the foundation for a new generation of AI systems capable of adaptive reasoning over large datasets, continuous learning, and efficient long-context processing, a capability that could influence future developments across both research and applied AI.
Read Next: Did Solana’s Top Lending Platform Mislead Users? Jupiter’s Risk Reversal Sparks Alarm