Anthropic Just Taught Claude to Dream

If you've ever been frustrated that a chatbot can't remember what you told it last week, this story is for you.

This week at Anthropic's Code with Claude developer conference, the company quietly rolled out a feature called "dreaming." The name is cute, but the idea behind it is one of the more practical advances in AI we've seen in months. Dreaming is a scheduled process that lets Claude agents review their past work, find patterns, and quietly improve themselves between sessions ¹.

In other words, Claude is starting to learn from experience the way the rest of us do. Slowly, in the background, while we're not looking.

What Dreaming Actually Does

Here's the problem dreaming is trying to solve. Most AI assistants today have what you might call a goldfish brain. They can hold a lot in mind during a single conversation, but the moment that conversation ends, the slate gets wiped. Tomorrow you have to re-explain your preferences, your project context, your file naming conventions, all of it.

The standard workaround is a process called compaction, where the AI takes a long conversation and tries to compress only the most important parts into its working memory ¹. That helps within one chat. It does nothing for the next one.

Dreaming works differently. It runs on a schedule, often when nobody is using the agent, and reviews everything that happened across past sessions. It looks for repeating mistakes the agent keeps making, workflows that several agents converge on, and preferences that pop up again and again across a team ². Then it cleans up the agent's memory, merging duplicates, resolving contradictions, and pruning stale notes.

Anthropic borrowed the idea from how the human brain works. While we sleep, our hippocampus replays the day's experiences and stabilizes what's worth keeping ³. Dreaming does roughly the same thing for AI agents. The brand name is whimsical. The mechanism is real.

Why This Matters Outside the AI Lab

If you're not building AI agents for a living, you might wonder why any of this is your problem.

Here's why. The single biggest complaint I hear from people trying to use AI at work, whether they're marketers, IT folks, or solo business owners, is that the AI keeps forgetting what they said. You teach it your tone of voice, you correct its formatting habits, you explain that no, the company name is spelled with a lowercase letter, and a week later you're explaining all of it again. That's exhausting, and it's a real reason a lot of people give up on AI tools after the novelty wears off.

Dreaming is the early version of a fix for that. It's the difference between an assistant who shows up green every morning and one who quietly remembers that you don't like Oxford commas and that the client in Phoenix prefers slide decks over PDFs.

The early results are interesting. Harvey, a legal AI company, said its task completion rates jumped roughly six times after adopting dreaming ³. That's a big number, and worth a healthy pinch of salt since it's the vendor's own measurement, but it points at the size of the problem. AI agents that can't remember what worked last time are leaving a lot of value on the table.

The Other Two Features Worth Knowing

Anthropic announced two more capabilities the same day, and they're easier to explain than dreaming.

The first is called outcomes. It lets you show an agent what a good final result looks like, not just describe it. So instead of writing a paragraph telling Claude what a great client email sounds like, you hand it three actual emails you've sent. A separate grader agent then checks the AI's work against your examples. Anthropic says this lifts task success rates by about ten points compared to plain text instructions ⁴.

The second is multi-agent orchestration. This lets a lead agent break a big job into smaller pieces and hand them off to specialist sub-agents. You can watch the whole tree of work in the Claude Console and see what each one did. Think of it as the AI equivalent of a project manager assigning slices of work to a team ¹.

There was also a quieter announcement that Pro and Max subscribers had their five hour usage limits doubled to ten ¹. That came thanks to extra compute Anthropic picked up by tapping into SpaceX's Colossus 1 supercomputer for inference ⁵. Heavy users have been hitting walls lately, so this one is a real quality of life win.

What This Means For You

For most working professionals, none of this requires action today. Dreaming is in research preview, which means developers have to apply for access and the rest of us are still using AI tools the old way.

But here's the trend worth watching. The frustrations you have with AI right now, the forgetfulness, the wandering quality, the need to babysit every output, those are exactly the problems the field is now actively solving. Dreaming is one piece of that. Outcomes is another. The agents you'll be working with a year from now are going to look more like attentive collaborators and less like talented amnesiacs.

That's a useful thing to keep in mind whether you're trying to figure out which AI tool to invest your learning time in or whether you're nervously eyeing a coworker who keeps demoing new automations. The technology is moving in the direction of remembering you, not replacing you. Pay attention to vendors who are leaning into that, and treat the ones who keep selling chat-only assistants in 2026 with a bit of healthy skepticism.

The AI that helps you most over the next few years won't be the smartest one. It'll be the one that finally bothers to learn who you are.