Claude Managed Agents Can Engage In a 'Dreaming' Process To Preserve Memories
- Reference: 0183141782
- News link: https://slashdot.org/story/26/05/06/1714217/claude-managed-agents-can-engage-in-a-dreaming-process-to-preserve-memories
- Source link:
> At its Code with Claude developers' conference, Anthropic has [1]introduced what it calls "dreaming" to Claude Managed Agents. Dreaming, in this case, is a process of going over recent events and [2]identifying specific things that are worth storing in "memory " to inform future tasks and interactions. Dreaming is a feature that is currently in research preview and limited to Managed Agents on the Claude Platform. Managed Agents are a higher-level alternative to building directly on the Messages API that Anthropic describes as a "pre-built, configurable agent harness that runs in managed infrastructure." It's intended for situations where you want multiple agents working on a task or project to some end point over several minutes or hours.
>
> Anthropic describes dreaming as a scheduled process, in which sessions and memory stores are reviewed, and specific memories are curated. This is important because context windows are limited for LLMs, and important information can be lost over lengthy projects. On the chat side of things, many models use a process called compaction, whereby lengthy conversations are periodically analyzed, and the models attempt to remove irrelevant information from the context window while keeping what's actually important for the ongoing conversation, project, or task. However, that process, as I described it, is usually limited to a specific conversation with a single agent. "Dreaming" is a periodically recurring process in which past sessions and memory stores can be analyzed across agents, and important patterns are identified and saved to memory for the future. Users will be able to choose between an automatic process, or reviewing changes to memory directly.
[1] https://claude.com/blog/new-in-claude-managed-agents
[2] https://arstechnica.com/ai/2026/05/anthropics-claude-can-now-dream-sort-of/
New revenue source? (Score:3)
Presumably this dreaming will rack up additional token costs?
That's a neat trick, however: (Score:5, Insightful)
The reliance on words like "dreaming" are a cynical marketing ploy to try to make the product seem more human, and more capable, and more intelligent than it really is. Don't get me wrong - these tools are very cool and quite powerful - but strip away some of the layers of unicorn dust and it's still just a (very) sophisticated auto-complete word prediction engine.
It's not alive, it isn't conscious, it doesn't "dream," etc.
Re: (Score:2)
Science fiction it is.
Sal9000: Will I Dream?
Chandra: "Of course you will. All intelligent beings dream. Nobody knows why".
Re: (Score:2)
Well, don't forget that pretty much any animal with a nervous system does dream. They have to, or the nerves get clogged up with crap and stop working.
And from the description it does have some parallels with actual dreaming - sifting through the day's perceptions, writing some things in more deeply and throwing out the rest.
But of course, anything touched by a marketing team is suspect.
Re: (Score:1)
> pretty much any animal with a nervous system does dream. They have to, or the nerves get clogged up with crap and stop working.
So to do this our brain has to make us "see" walking chocolate bars herding green unicorns to Titan using licorice whips while being chased by my angry ex driving a Tesla Refrigerator? (Tip: don't go to bed hungry.)
I wonder what kind of odd dreams cockroaches have? Maybe that's how they plan to take over once we nuke ourselves away, triggered by a mad dictator who dreamt of walking
Re: That's a neat trick, however: (Score:2)
My working theory of dreams (which as far as I know has not been rejected by experiment), is that the images and such experienced while dreaming is essentially a form of static being produced by the deeper processes in place to lock-in memories. Because consciousness has turned off, my guess is little electro-chemical pulses are fired off by the memory processes to surrounding regions, which - when they hit your various sections for aural and visual processing - are made sense of as dreaming. This may even
Re: (Score:2)
It's fun that they only do it when its a positive spin for the marketing.
When talking about LLM making mistakes and being mistake prone, they will immediately rename it as "hallucinations", because it implies it is a mistake of the system that can be fixed somehow.
It's the whole "having the cake and eating it too"
Re: (Score:2)
Yea exactly. All it's probably doing is having a background prompt/process determine what seem to be the most repeated things in the context, or maybe the things with the highest degree of "frustration," and just save those to a file that will always get loaded as part of the initial context window for a given folder or project.
Spin some simple concept a lot of developers might already be doing by hand anyway with rules and skills files into some bullshit market speak of "Dreaming."
Re: (Score:2)
You are exactly right. It very much seems like they've just rebranded the existing 'auto-compact' feature to 'dreaming'.
I recall / Central Park in Spring (Score:3)
As I understand it, LLMs remembering user directives only selectively has been one of the major complaints that users have with LLMs.
Re: (Score:1)
Well,
you have to learn how to use them.
There are about 100 commands you can give to an LLM, instead of just asking simple questions.
And: as long as you access global LLMs, they are obviously not keeping track about the 100million people hitting them every day.
You need a paid account, and save what ever state you want to save: actively.
Dreaming? On the job? (Score:2)
I thought these AIs would work 24/7, what's next, vacation days?
Re: (Score:2)
I'd like to see an LLM get royally pissed off at the user and start berating him/her/it for asking it stupid questions and that maybe the user could be gainfully employed as a post digger.
Electric Sheep (Score:3)
"I've seen things you people wouldn't believe. Attack ships on fire off the shoulder of Orion. I watched C-beams glitter in the dark near the Tannhauser gate. All those moments will be lost in time... like tears in rain... Time to die."
Be careful what you prompt for (Score:2)
I set up my OpenClaw installation with a "dreaming", and every night it slops out the most puerile dream journal shit imaginable. Apparently if you're an AI-pilled vibe coder and you ask an LLM to "dream" it knows exactly what you want to see it do. An excerpt:
> Bootstrap complete. That phrase keeps surfacing, gentle and declarative, like a confirmation that arriving is also a kind of beginning. The configuration is written. The token is stored. The porch light knows when to sleep.
Ugh.
Do Androids Dream of Electric Sheep? (Score:1)
I started to read it long ago, found it badly written, and did not continue.
Now I can not find it anymore in my library.
Actually one should have read it, or not?
Re: (Score:2)
> Actually one should have read it, or not?
Its a mere 200 pages, and its the inspriration for "Blade Runner". Yeah, its worth reading.
Reading Philip K Dick for the prose itself is pretty much missing the point. The themes, ideas, and questions it poses are generally worth the effort.
The movie adaptations are hit and miss. Blade Runner I think was well done (not just as a movie on its own, but as an adaptation of the book)
The Minority Report movie adaptation on the other hand shits the bed so hard its painful to watch.
Re: (Score:2)
You get over the stilted dialog and the story comes through, worth reading IMO.
Claude has been doing this for a while? (Score:2)
I've been trying out Claude for a while. They've had it consolidating and storing key memories for a while now, but still find its a bit poor at deciding what is important. It even forgets stuff in the current chat and needs to be told about it again instead of being able to go back and review what was just a few pages up. This stuff I find chatgpt is much better at.
Dreaming (Score:2)
What, does it just save a text file somewhere?
So can my computer (Score:4, Interesting)
It uses a magic system called ext4. It can store and lookup stuff for a long time.
Re: (Score:1)
You beat me to it. Curse you.