Anthropic shows how Claude 'dreaming' fixes agent memory

What Anthropic published

On November 6, 2026, Anthropic published The components of agent memory ↗, an engineering post by Lance Martin. It covers how Claude decides what to remember between sessions and how an offline process Anthropic calls "dreaming" cleans that memory up. Anthropic's post on X ↗ about it passed 13,000 likes and 1.7 million views within two days. Interesting Engineering ↗ covered it on November 7.

Dreaming itself isn't new. Anthropic added it to Claude Managed Agents as a research preview in May, as The Decoder reported then ↗. What's new is the evidence: Anthropic's first detailed numbers on what each part of memory actually does.

The idea borrows from sleep. Agents write notes during a session, based on what seems important at the time. Some of those notes are wrong, and others go stale. Dreaming runs later, rereads the original session transcripts, and fixes the notes against what actually happened. It can also build linked "wiki" pages and choose a small set of notes to load at the start of every session.

The numbers Anthropic shared

These are Anthropic's own evaluations, not independent tests:

• Less memory, more progress. In Pokémon Red, about 10,000 steps in, Claude Sonnet 3.7 had written 413 memory files across 225 folders and earned one badge. Claude Fable 5 kept nine files in two folders plus one global.md and earned four badges.

• A corrected note changed the outcome. In one run Claude wrote down the wrong side of Mt. Moon for a PokéCenter and walked off a one-way ledge three times. Dreaming checked the note against the transcript and rewrote it, and that run reached Cerulean City. The run with the uncorrected note never got past the mountain.

• Wikis grow with dreaming. On roughly 350,000 internal files (~420M tokens) from Stainless, the company Anthropic acquired in May, Opus 5 agents built a wiki over successive dreaming sessions. Answers from the wiki alone rose from 13% to 48% of graded criteria as dreaming scaled from 2 to 128 sessions. Wiki plus search reached 61%, against 53% for search alone, and saved about 20,000 tokens per question.

• Instructions matter. On a Harvey LAB legal-brief task, a wiki built with generic instructions kept 41 of 57 contract facts. After Claude revised its own instructions (one agent per contract, quoting terms word for word), it kept 57 of 57.

• Don't over-steer strong models. Telling Opus 5 and Opus 4.7 to "record only failures" cost them much of the gain memory gave them. Older Sonnet 4.6 nearly doubled its score when prompted to read its memory first and test its theories.

Anthropic is also open about where it falls short. On one legal task the wiki showed no gain over plain search. On another, the wiki grew larger than the source documents themselves.

Who can use it

Developers can use it, though most chat users can't yet. Per the Dreams documentation ↗, Dreams is a research preview on Claude Managed Agents, and you have to request access. A dream is an asynchronous job. It takes one memory store and 1 to 100 past sessions, then writes a new store. The input is never modified, so you can review the output and throw it away. Jobs take "minutes to a few hours," are billed at standard token rates for the model you pick, and aren't eligible for zero data retention. Supported models include Opus 5.5, Sonnet 5.5, Fable 5.1 and several older ones.

The post also describes how Claude Code's auto memory keeps its always-loaded layer small, loading only the first 200 lines or 25 KB of MEMORY.md.

What people are saying

Builders mostly took it as a pattern to copy. Developer superalesha wrote:

"Anthropic just showed how Claude "dreams". I am stealing it for my 3090s."

— @superalesha ↗

Some pushed back on how it was being sold. "This is not an announcement that every consumer Claude chat does it," noted @thataibloke ↗. Others described their own versions. One developer said their team's markdown memory vault now has "a separate reconciler agent" that "dreams nightly," adding: "Lossy, but each forget is a commit you can undo" (@Tel_riker ↗). And one summary put the lesson in a line: "Memory without a replay is just confident fiction" (@Tan07020 ↗).

Our take: who should care

If you build long-running agents, this is the most useful thing Anthropic has published this month. It's a design guide more than a product launch, and its lessons apply to any stack:

1. Store what the agent can't see again. Strategy, tested rules and known traps belong in memory. Facts the environment already shows don't.

2. Validate memory against the source. A wrong note that gets loaded every session is worse than no note. Schedule a cleanup pass that checks notes against transcripts or logs.

3. Keep the always-loaded layer small and link out to the details.

4. Wikis help on multi-document questions, not single-document lookups. Measure before paying the compute.

Should you request Dreams access? It's worth it if you already run Claude Managed Agents with memory stores and see stale notes piling up. Because the output is a separate store, the experiment is low-risk. If you're on another framework, the same loop is easy to build yourself; see our AI agent frameworks comparison and MCP guide. Meta's Active Memory push for Muse shows the consumer side of the same race. And if you live in Claude Code, its MEMORY.md limits now come with a published reason (our Claude Code review).

Sources

• Anthropic: The components of agent memory ↗, November 6, 2026

• Claude Platform docs: Dreams (research preview) ↗, checked November 8, 2026

• Interesting Engineering: Can Claude dream now? ↗, November 7, 2026

• The Decoder: Claude's new "Dreaming" feature ↗, May 7, 2026