[Digital Today reporter Jinju Hong] Google has released research on a so-called “sleep framework” that allows large language models (LLMs) to keep acquiring new knowledge even after they complete training. The core idea is to separate the process of taking in new information from the process of organising existing memories, so the AI can update its knowledge without forgetting what it previously learned.

On Oct. 30 local time, online media outlet Gigazine introduced a new framework that Google recently presented to reduce knowledge loss during continuous learning in LLMs. The research was released through a research post titled “A New Way to Preserve Knowledge in LLM Continual Learning” and a paper titled “Language Models Need Sleep: Self-Repair and Memory Consolidation Learning.”

LLMs such as Gemini and GPT are formally released after learning vast amounts of data and undergoing performance testing. The issue is that a time gap arises between when training ends and when users actually use the model. As a result, a released model may not fully reflect the latest information.

According to the outlet, Google’s Gemini 4 Argon, which appeared in September 2026, was also trained on a dataset through March 2026. Events that occur after training, or newly accumulated information, are supplemented by using external functions such as web search.

Even so, continuously injecting new data into a model is not a simple solution. That is because it can cause “catastrophic forgetting,” in which previously learned knowledge disappears as the model takes in new information. An approach separate from continuous learning is needed because attempts to reflect the latest information can degrade existing knowledge or performance.

The sleep framework proposed by Google divides how the model operates into a “wake phase” and a “sleep phase” to address this problem. In the wake phase, the model processes new inputs and stores information in a short-term memory component. In the sleep phase, it stops processing new inputs and consolidates the stored information while organising and improving the model’s internal memory.

The structure is inspired by the way people take in new experiences while awake and organise memories while asleep. Rather than having the AI accumulate new information indiscriminately, it is designed to organise existing knowledge and newly acquired information in a separate phase.

Google said this makes it possible to introduce new information while preserving existing knowledge. It also suggested the possibility of reducing catastrophic forgetting during continuous learning and enabling models to keep accumulating knowledge even after training ends.

The research could also change how AI models ensure they remain up to date. Until now, models that completed training have often relied on web search or external databases to reflect information that emerges after release. The sleep framework is an approach that aims to complement such limits by absorbing new information into the model itself and integrating it with existing memory.

Still, additional verification is needed before the research-stage technology can be applied to commercial AI services. Key tasks remain, including how accurately it reflects new knowledge and whether it can sustain long-term learning while maintaining existing performance. In future AI competition, how efficiently knowledge can be accumulated and retained after release, as well as the scale of initial training, is expected to be an important differentiator.