Anthropic Lets AI Dream

Dear curious, creative, explorers,

San Francisco, Wednesday, 6 May 2026, somewhere between the keynotes. Simon Willison sits in the room, live-blogging, and types one unassuming sentence: „Dreaming looks really interesting.“ On stage, a feature has just dropped that isn’t called „GPT-something“, doesn’t add ten billion parameters, doesn’t smash a new benchmark. It is simply called Dreaming – and that’s exactly my AI Gem of the week.

„All men dream: but not equally. Those who dream by night in the dusty recesses of their minds wake in the day to find that it was vanity: but the dreamers of the day are dangerous men, for they may act their dreams with open eyes, to make it possible.“ — T. E. Lawrence, Seven Pillars of Wisdom (1922)

What’s it about? AI agents that go to sleep between sessions, comb through their own transcripts, lift patterns, identify mistakes – and wake up the next morning a little smarter. Sounds like science fiction. As of last week, it’s an API.

Dreaming: When AI agents work on themselves between shifts

What Anthropic unveiled on 6 May at its developer conference Code with Claude as a Research Preview sounds harmless at first: a new feature for so-called Managed Agents. Read the blog post slowly and the impression shifts. Dreaming is a scheduled, asynchronous process. You hand your agent an existing memory store, optionally up to 100 past sessions – and it reconstructs a new, cleaned-up store from them. Duplicates merged, contradictory entries updated, new insights surfaced. The whole thing runs while you sip your coffee (or sleep – fittingly).

Sounds unspectacular? It isn’t. So far, every large language model wrestles with the same structural problem: it has no long-term memory. Whatever doesn’t fit into the context window is gone. „Compaction“ – squeezing conversations together – works only within a session. Dreaming takes the next step and works between sessions, even across multiple agents on the same team. Crucially (and for me the real charm here): no model weights are touched. The agent writes its lessons as plain-text notes and structured „playbooks“ – inspectable, auditable, correctable. If you’re skeptical, you can review the output before it overwrites anything; the technical docs spell that out clearly.

The early numbers are interesting. Legal AI company Harvey reports a roughly sixfold increase in completion rate for long document workflows since their agents started „dreaming“ between sessions. Netflix uses the sibling feature Multiagent Orchestration to comb through hundreds of build logs in parallel. Wisedocs cuts document review in half. Sounds like the usual vendor anecdotes – yet it points in a direction worth remembering.

The real thing happening here: agents learn between deployments. Not through expensive re-training, but through reflective notes.

A small shift. But one that could permanently change the relationship between model and memory.

How to picture this in practice? We all know it: you’re sitting with a new tool, explaining for the twelfth time the same edge case in your Excel export, and sighing. Imagine the agent remembering this edge case – not for the next reply, but for the next week, the next quarter, the next colleague. That’s the step that’s been on my mind this week. We’ve thought about this once before – in the KI-Logbuch issue „Alleingang? Wie sich KI durch selbstverbessernde Algorithmen selber optimiert“ (published 5 March 2025). Back then it was about models training themselves. Today it’s about something more humble – and perhaps more practical: agents taking notes for themselves.

Outcomes, multi-agent orchestration – and the biological cousin

Dreaming didn’t come alone. Anthropic moved two other features into public beta in the same breath: Outcomes and multi-agent orchestration. Outcomes is, at heart, a rubric. You describe what „good“ looks like – a structure, a brand voice, a list of must-haves – and a separate grader agent checks in its own context window whether the result delivers. If not, it bounces back to the main agent. In Anthropic’s tests, as documented by VentureBeat, that lifts task success by up to ten points, with the biggest gains in docx and pptx generation. Multi-agent orchestration in turn lets a lead agent break complex jobs into pieces and delegate them to sub-agents – each with its own model, prompt, tools. If you want to watch this in real time, The New Stack has a fine overview of the three building blocks. Three levers, one goal: agents that need less hand-holding on complex tasks.

What neuroscience has known for decades

The „dreaming“ metaphor isn’t a PR trick. It points to a well-documented insight in neuroscience since the 1980s: during sleep, especially deep sleep, the hippocampus „replays“ experiences from the day in compressed form – so-called memory replay in coordination with sharp wave ripples – and „teaches“ them to the neocortex. That’s how memories migrate from short- to long-term storage. A review in Neuron describes this as a coordinated interplay of thalamic spindles, cortical slow oscillations, and noradrenergic activity. Anthropic’s „dreaming“ is not a one-to-one transfer of that mechanism – but the functional analogy is uncannily precise: sift experience, compress it, consolidate, keep what matters, discard the rest. And by the way: as early as 2023, Danijar Hafner showed with DreamerV3 that „dreaming“ – in the sense of imagined rollouts in a world model – can beat existing reinforcement learning methods across 150-plus tasks. The metaphor has been circling for years. It’s now arriving in the enterprise stack.

Did you know?

The term Dreaming didn’t appear in AI research only last week. As early as 2023, Danijar Hafner and his team published DreamerV3 – a reinforcement learning algorithm that dreams its next steps inside an internal world model before acting. The system, by the way, was the first to find diamonds in Minecraft autonomously, without human examples. Dreaming as a term is therefore less a marketing trick than a nod to an entire line of research – a hint that the giants here are standing on shoulders.

Tips and tricks – if you want to experiment with agent memory yourself

Use Obsidian together with a strong AI model. You can already try a lot on a small scale here and let AI work (or dream?) with your notes.
Start small. Don’t throw a hundred sessions into a dream job at once. Three to five are enough to see what your agent recognizes as a „pattern“ – and what’s noise.
Review mode first. Have the newly curated memories shown to you before the agent adopts them automatically. Anyone who blindly trusts the AI here builds bias into the pipeline.
Clean playbook names. When your agent writes „decent-playbook.md“ or „pptx-quirks.md“, a clear naming convention helps – or in three weeks you’ll have a memory store full of cryptic fragments.
Measure first, dream later. Define a small baseline (e.g. ten standard tasks) before you enable dreaming. Otherwise you can’t tell later whether the „sixfold improvement“ also applies to you – or whether you’re just convincing yourself.
Abstinence is a move, too. Not every workflow needs memory. One-off tasks do fine with a good prompt. Memory is care – and care costs time.

What remains?

What remains? For me, above all this: the tech world has been talking for months about „bigger models“, „more data centres“, „more GPUs“. Last week Anthropic did something different – it showed a small, almost boring, almost meditative feature. One that does not bet on more parameters but on more reflection. That is – hand on heart – unusual for this industry. And perhaps exactly what we need right now. AI that doesn’t get loud, but learns quietly. That doesn’t want to talk, but to tidy up. That doesn’t produce gloss, but clarity. Will it last? No idea. Research Preview means: could become good, could fizzle out. But I like the direction. I’ll try it, I’ll report back – and I’d love to know: where would you let your agents dream? In which workflow would memory be worth more than ten more prompts?

„Who looks outside, dreams; who looks inside, awakes.“ — Carl Gustav Jung, Aion (1951)

Thanks for thinking along this week. I look forward to your comments, your examples, your skepticism – and to your dreams in the pipeline.

Until next week,

Yours, Arno