Google Research has proposed a 'sleep' framework that would let large language models keep learning after deployment, moving new information from short-term context into long-term parameters during scheduled offline periods. The approach, described in a blog post by Ali Behrouz and Vahab Mirrokni of Google Research, targets a long-standing problem. Today's models are largely static after training, with a fixed knowledge cutoff, and simply fine-tuning them on new data tends to cause catastrophic forgetting, where performance on earlier tasks degrades. The framework alternates between an active 'wake' phase, in which the model takes in inputs and stores temporary information in fast memory, and a 'sleep' phase, in which it pauses new input and runs two processes. The first, memory consolidation, expands capacity using Mixture-of-Experts modules and distills short-term context into those newly added parameters, after which the temporary parameters are reset. The second, 'dreaming', has the model generate synthetic training examples, randomly activates experts to mix knowledge domains, filters the examples using gradient-based selection, and fine-tunes on the highest-rated samples. The authors say the approach lets models absorb new facts, retain long-context knowledge up to 10 million tokens, and improve complex reasoning, while keeping the underlying architecture unchanged. The post describes the method and its reported results. It does not announce a product release or independent evaluation, and its comparisons to neuroscience are analogies rather than biological claims. For Australian businesses and professionals, this changes nothing today. It is a research proposal with no released tool to buy or trial. If similar methods reach commercial models, the main considerations would be the cost of extra training during sleep cycles, and the risk that models retain new information, including personal information, inside their weights. That would raise questions under the Privacy Act 1988 about collection, use and deletion, and about whether vendors can reliably remove or correct that information. Until then, it is worth tracking rather than acting on.