IT News

Artificial intelligence faces one of its greatest challenges: achieving continuous learning and retaining information long-term, similar to the functioning of the human brain. To address this limitation, Google Research has introduced Nested Learning, a new machine learning paradigm inspired by the dynamics of brain waves.

This approach seeks to provide AI models with a more robust memory and the ability to incorporate knowledge in a sustained manner. The proposal was presented at the 39th Conference on Neural Information Processing Systems (NeurIPS 2025) and could mark a before and after in the development of more adaptive and efficient AI systems.

If you have ever interacted with an Artificial Intelligence (AI) model, such as an advanced chatbot, you will have realized that, while it is incredibly intelligent in answering your questions at the moment, it has a terrible memory. If you close the conversation and return the next day, it will have "forgotten" much of what you talked about.

This is not just a software problem; it is the biggest challenge in modern AI: continuous learning. Current models are great at what they were trained for, but they struggle to learn new things after their initial training. Google Research has tackled this problem at its root, looking at the best expert on the subject: the human brain.

The Problem of "Artificial Amnesia" in AI

To understand why Nested Learning is such a great advance, we must first understand the great limitation of current AI.

Despite advances in language models and deep architectures, current AI remains essentially static after its initial training. Large Language Models can only execute tasks learned during the pre-training phase, without continuously incorporating new skills or knowledge.

When a Large Language Model (LLM), such as Gemini, is trained, it is given a massive amount of data and its artificial brain (its parameters) "crystallizes." After that, it can only operate with the information it already has.

Google Research explains that this limitation is comparable to anterograde amnesia in humans. Think of a person who, due to an injury, can only remember what happened up to the time of the trauma, but cannot form new long-term memories. Analogously, AI:

• Can only operate with the information in its context window (short-term memory, what we discussed in the last few minutes).
• Lacks the ability to consolidate that new information into its long-term parameters.

If an LLM learns a new piece of data, it has two options: either it forgets it, or it completely rewrites all its knowledge, which is inefficient and leads to what experts call "catastrophic forgetting." This leads us to the key question: how does the human brain manage to learn without erasing what it already knows?

The Biological Inspiration: Fast Memory and Slow Memory

The human brain is a master of continuous learning, thanks to neuroplasticity and a two-speed memory consolidation system:

1. Fast Consolidation (Online): This happens immediately after you experience or learn something. Your brain quickly stabilizes it. It is "working memory" or short-term memory.
2. Slow Consolidation (Offline): This is what happens, mainly, while you sleep. Brain waves of different frequencies interact to reorganize and reinforce memory, moving knowledge from temporary storage areas (like the hippocampus) to permanent storage areas (the cortex). This is where new knowledge is integrated without erasing the old.

In current AI models, these "two-speed memory" mechanisms do not exist. If they learn, they do so with their entire system at the same time, which causes the instability we discussed.

The Nested Learning Solution: Nested Optimization

This is where the genius of Google Research comes in. Nested Learning proposes an architecture that mimics this hierarchy of biological memory.

Instead of having a single artificial brain that learns at a single speed, Nested Learning organizes the model as a system of nested optimization problems.

What does "Nested" mean?

• Imagine several layers of neural networks, like Russian nesting dolls (matryoshkas), where each layer represents a different temporal scale and level of abstraction.
• High-Frequency Layers (Close to the Surface): They are constantly updated and handle the smallest details and immediate data (short-term memory, or the equivalent of "fast" sleep or online consolidation).
• Low-Frequency Layers (Deep): They are updated slowly and store fundamental concepts and stable knowledge (long-term memory, or the equivalent of offline consolidation during deep sleep).

This structure allows the AI to do the following:

1. Process information with different degrees of abstraction: The superficial layer processes the exact words, while the deep layer consolidates the general concept.
2. Integrate new knowledge without forgetting the old: New data is first tested and stabilized in the superficial layers. If they are considered relevant, only then are they integrated into the parameters of the deeper layers, reorganizing the memory in an orderly fashion, instead of erasing everything.

The development of Nested Learning at Google Research involves multidisciplinary teams seeking to address one of the greatest limitations of current artificial intelligence. (Image Illustrative Infobae)

HOPE: The Demonstration Module and the Results

To demonstrate that Nested Learning works in practice, researchers developed a module called HOPE (an acronym), an auto-referential architecture that uses this nested learning along with a continuous memory system.

In tests conducted and presented at NeurIPS 2025, HOPE consistently outperformed well-known traditional models (such as Transformers and DeltaNet). Specifically, it showed:

• Lower perplexity levels: This is a technical metric that indicates how "surprised" or uncertain the model is by new information. Lower perplexity means the model handles new information with greater confidence.
• Higher accuracy rates: The model was more accurate in language and continuous reasoning benchmarks.

This suggests that HOPE, thanks to Nested Learning, is more capable of adapting and learning after its initial training, without falling into artificial amnesia.

This paradigm overcomes the traditional dichotomy between short-term and long-term memory present in previous models, allowing artificial intelligence to manage data abstraction and consolidation in a more dynamic and precise way. (Image Illustrative Infobae)

The Future of AI is Neuroscience

The introduction of Nested Learning is a milestone because it marks a paradigm shift. We are no longer trying to make AI just bigger or faster; we are trying to make it more human in its functioning.

By basing the memory architecture of AI on the principles of neuroscience (brain waves, online and offline consolidation), Google Research not only improves the efficiency and flexibility of artificial memory. It also opens the door to the creation of much more expressive, adaptive, and reliable systems.

If future AI models adopt Nested Learning, you will be able to have virtual assistants that truly know you and remember your preferences long-term without needing to be constantly retrained. This is the path for Artificial Intelligence to become a truly continuous and organic tool in our daily lives.

Nested Learning allows computational models composed of multiple levels to process data with different degrees of abstraction and temporal scales, bringing artificial intelligence closer to the learning and memory capacity of the human brain, according to Google Research.