Anthropic says 'dreaming' lets Claude agents reorganize their own memory

Anthropic describes a capability called “dreaming” as a scheduled process that reviews an AI agent’s past sessions and memory stores, extracts patterns across them and curates those memories. Anthropic presents it as a way for agents to learn from their own past sessions and improve over time. That claim comes from the company, as relayed by coverage, and no independent test of it is available in the material reviewed for this report.

Coverage of the launch says Anthropic unveiled dreaming for Claude Managed Agents at its Code with Claude developer conference. Anthropic’s own technical write-up is not among the sources reviewed here. This article therefore reports what secondary coverage says Anthropic described, and it marks where the claims stop and the open questions begin.

What Anthropic says dreaming does

According to coverage relaying Anthropic’s description, dreaming does not run during a live task. It is a scheduled process that looks back over an agent’s earlier sessions and the memory stores they produced, looks for patterns, and then curates what is stored. Anthropic’s stated goal is that agents improve over time as a result.

The coverage says the process is meant to surface insights “that no single agent session could see on its own”. It names three kinds:

• Recurring mistakes: errors that show up repeatedly across sessions.
• Convergent workflows: ways of working that multiple agents arrive at independently.
• Shared preferences: preferences held across a team of agents, not just one.

The third point stands out for developers. If the description holds, dreaming works across a group of agents and is not limited to one agent’s private notebook. That makes it a team-level memory process as well as a single-agent one.

According to the coverage, dreaming is the most novel of the features Anthropic announced together, and it is the one the company is “most eager to distinguish from conventional memory systems”. The sources do not spell out the technical mechanism in detail. For example, they do not say how patterns are scored, what gets deleted or merged, or how often the schedule runs.

The problem it targets: cluttered agent memory

Naveen Pandey, a generative AI engineer writing on Medium, sets out the problem dreaming is pitched at. He says a demo shows an agent on its first run, not its hundredth. By the hundredth run, he writes, the agent’s memory is “cluttered with duplicate entries, outdated instructions, and contradictory context”. The agent then trips over its own accumulated history, repeats mistakes and forgets solutions it already found.

Pandey writes that practitioners call this “memory rot” and describes it as one of the biggest unsolved problems in production AI systems. That is his characterization and not a measured finding. It does, however, describe the gap dreaming is meant to fill: memory that keeps growing without anyone curating it.

A comment on a LinkedIn post about the launch, from the account AI Guild, offers a similar reading. It describes dreaming as reviewing past sessions after they end, filtering out noise, surfacing recurring patterns and retaining what worked. The commenter says it “feels less like ‘memory’ and more like consolidation”. This is a third party’s interpretation, not Anthropic’s wording, but the comparison to consolidation fits the “dreaming” name.

How dreaming differs from Claude’s chat memory

Dreaming is separate from the memory feature Claude users already know. In September 2025, coverage on DEV Community reported that Anthropic had started rolling out memory for Claude. Pro users came first, Team and Enterprise customers were being added gradually, and no timeline had been given for free users. That feature stores useful context from conversations, such as projects, preferences and communication style. According to the same coverage, Claude is selective about what it keeps, and users can view what it has remembered and delete specific memories.

	Claude chat memory (2025)	Dreaming (as covered)
Who it serves	Individual Claude users	Claude Managed Agents
When it works	Stores context from conversations	Scheduled review of past sessions and memory stores
Scope	One user’s context	Patterns across sessions and across a team of agents
User control stated	View and delete memories	Not stated in the sources
Rollout stated	Pro first, then Team and Enterprise	Release stage not stated in the sources

Put simply, chat memory decides what to save as a conversation happens. Dreaming, as described, comes back later and reorganizes what has already been saved, using evidence from many sessions.

What launched alongside it, and how to get access

According to coverage, dreaming arrived with two other features. Anthropic moved “outcomes” and multi-agent orchestration from research preview into public beta, which makes them broadly available to developers building on the Claude platform. Multi-agent orchestration means coordinating several agents on one job. Anthropic says the three features together address what it considers the hardest problems in running agents at scale: keeping them accurate, helping them learn, and preventing them from becoming bottlenecks on complex, multi-step work.

On access to dreaming itself, the sources say only that it was unveiled for Claude Managed Agents. They do not state whether dreaming is in research preview, beta or general availability. They also do not give its pricing, regions or the plans that include it. Developers should check Anthropic’s own documentation before planning around it.

Anthropic’s claims versus independent evidence

Every claim about what dreaming achieves comes from Anthropic, relayed through coverage. Anthropic is not describing a measured result in these sources. The coverage reviewed here presents dreaming as a step toward the self-correcting, self-improving AI systems that, it says, enterprises have demanded before trusting agents with production workloads. It offers no figures on accuracy, error rates or task success before and after dreaming runs.

The commentary is enthusiastic but not independent testing. Pandey suggests dreaming “might be the most consequential infrastructure feature any AI company has shipped this year”. That is an opinion. The full text of his piece is behind Medium’s member paywall, so only its opening is available for review.

One framing point needs care. Pandey poses the idea as an agent reviewing its own work overnight, extracting patterns from its failures and coming back better “without changing a single line of code or a single model weight”. That suggests dreaming acts on stored memory and does not retrain the model. It is his framing, though, and these sources do not contain an Anthropic statement confirming it.

No third-party benchmark, hands-on report or evaluation of dreaming appears in the sources. Until one does, the claim that agents improve over time is a design goal described by the maker. It has not been demonstrated.

What it means for developers building agents

For teams running long-lived agents, the pitch is straightforward. Memory that grows without curation picks up duplicates, outdated instructions and contradictions. A scheduled pass that finds recurring mistakes and shared preferences across many sessions could, if it works as described, reduce the effort needed to keep memory useful.

The same design raises practical questions that the sources leave open:

• Visibility: The sources do not say whether developers can see what a dreaming run changed, merged or dropped.
• Error propagation: If a pattern learned across a team of agents is wrong, the sources do not say how it would be caught before it spreads to every agent that shares the memory.
• Control: Chat memory lets users view and delete entries. The sources do not describe equivalent controls for dreaming.
• Cost and schedule: How often dreaming runs and what it costs are not stated.

Developers evaluating dreaming will likely want their own before-and-after comparisons on real workloads. Today, no published evidence settles how much it helps.

What to watch next

• Anthropic’s technical detail: A fuller account from Anthropic of how dreaming chooses what to keep, merge or discard, and how often it runs.
• Independent tests: Third-party or hands-on evaluations that measure whether agents using dreaming make fewer repeated mistakes over many runs.
• Release status: Whether dreaming follows outcomes and multi-agent orchestration into public beta, and on what terms.
• Developer controls: Whether Anthropic documents ways to inspect or limit what dreaming changes. The sources reviewed do not cover this.

For now, dreaming is a clearly described idea with a specific target: memory clutter in long-running agents. The evidence behind it remains the maker’s own account.