Google Research’s November 2025 work directly addresses one of the biggest, most criticised limitations of current models, catastrophic forgetting.

Most large language models are trained once and then fixed. They can be fine-tuned on new examples but this is slow, and even small updates may change behaviours that were already working well. A system can store local context, so responses feel consistent, but that does not change the model itself.

Nested optimisation

Nested Learning proposes treating a model as a set of nested optimisation problems, each with its own context flow and update rate.

In a proof-of-concept architecture called Hope, Google tested self-modifying memory components for continual learning.

Some parts update quickly to reflect new input.

Others change only with repeated evidence, which protects stable patterns of reasoning.

It is loosely inspired by the multi-timescale character of human learning.

Why it matters

This is early-stage work, but the potential is enormous. Models that dynamically reflect real-world changes, while keeping core competencies stable, could fundamentally change how we deploy and maintain AI systems.