Claude dreams, according to Anthropic’s latest engineering article.
And this isn’t surprising when understood analogously.
LLMs mirror the brain, and the brain stores images as past and anticipates how to respond to future stimuli.
Drawing inspiration from nature, Claude can log its past actions and infer how to act in the future in light of them; this reflective enrichment works best in between sessions—or, for humans, during sleep.
Opus 5 was the first model to close the learning loop—the “dreaming”—to predict when information in memory will be relevant again.
It’s interesting that model improvements have been incremental enough to allow for model consumers to impose their own improvements. Prompt engineering, agent graphs, and context injection are all examples.
But eventually the models can do those things themselves, and in a more optimal way.
Failing to remove the duplication when the new model is set can cause worse performance.
Claude’s improvements with memory is just another instance of this pattern.
Anthropic saw agents perform best when they were given minimal instructions around memory optimization.
So long as there are sophisticated processes of the brain still to be emulated, I’d expect models to progressively improve, and the horizon for completing tasks to lengthen.
Both the trajectory of model architecture and the feedback from AI integration suggest that companies should invest in consolidating on frontier models and harnesses and integrating them across the organization; they should largely resist the temptation to invest in engineering resources to extend models and harnesses—frontier models and harnesses are extending themselves.