Hopp til hovedinnhold

Anthropic: Claude agents can “dream” to organize memory between sessions

Anthropic details agent memory and “dreaming”: idle time to fix mistakes, build wikis and promote notes. Engineering post 6 Nov — not a new model launch.

Av Joachim HøgbyKilde: Anthropic

Anthropic explains how Claude agents can “dream”: use idle time between sessions to organize memory, correct mistakes, and build representations future sessions can trust. It is an engineering blog post about agent memory — not a new model or product launch.

The post went out via Anthropic’s engineering blog and the official @AnthropicAI account on Friday 6 November 2026 (22:16 Europe/Oslo on X). “The components of agent memory” lays out how the company thinks about memory for agents, and what role “dreaming” plays in making that memory more reliable over time.

What Anthropic actually describes

According to Anthropic, agent memory is knowledge Claude gains from experience — interactions and assigned tasks — and retains between sessions, typically as files or notes it can read later. It excludes weight updates, pure in-context learning inside one conversation, and compaction of long threads.

The company draws a parallel to human memory: fast encoding in the moment, and consolidation when the brain is “offline.” For agents they propose a similar division of labor. Agents log their work; idle time between interactions can reorganize context — an idea Anthropic links to “sleep-time compute” from Letta and UC Berkeley. That background process is what they call dreaming.

The post walks through several components and evaluations:

  • Memory notes as predictions. Writing down a rule is a bet that it will matter again. Claude can tag salient events in a session log or write them into a memory store.
  • CL-Bench 1.0. A long-horizon memory benchmark where Claude starts each question in a fresh session but can retrieve files from a persistent memory directory. Anthropic describes progress from scattered untested theories (older Sonnet) to closed learning loops on Opus 5 / Opus 5.5 and Sonnet 5.5, where memory captures validated rules and traps.
  • Pokémon Red as a salience test. Sonnet 3.7 wrote hundreds of files about every battle and item and earned one badge; a newer model in the text (Fable 5) wrote nine files plus global.md and earned four badges — by remembering strategy and rules the game state does not already show.
  • Dreaming can correct mistakes. In one Mt. Moon scenario Claude had recorded the wrong PokéCenter location, looped, and gave up. After dreaming checked the note against the transcript, Claude fixed the memory and made it through the cave.
  • Dreaming can build wikis. Across 128 sessions on a large internal corpus (Stainless, ~350,000 files), performance on 25 questions rose from 13% to 48% of graded criteria as Claude built and maintained a wiki. Wiki plus search added further gains and saved about 20,000 tokens per question in Anthropic’s measurement.
  • Hierarchy and injection. Dreaming can promote short indexes to the top of the memory hierarchy so the harness injects them into every new session — with a warning that overly long injections reduce adherence (same logic as Claude Code’s auto-memory cap of 200 lines / 25 KB of MEMORY.md).

What Anthropic does not claim

This is not a launch of a new Claude model, not a consumer feature that “turns on dreaming for everyone,” and not an enterprise package with an SLA. The numbers and Pokémon/wiki experiments are Anthropic’s own engineering evaluations. The post explicitly says many questions remain: when Claude should write memory mid-session versus in the background, how to keep memory fresh without too much compute, and when plain transcript search is enough.

Keep this separate from OpenAI’s ChatGPT “dreaming” (product memory for chat, covered earlier) and from Meta’s Active Memory / Active Focus roll-out to Muse and Meta AI the same day. The primary source here is Anthropic’s engineering explanation of agent-memory design.

What this means for leaders

  • Agent memory is an architecture decision. When vendors say agents “remember,” ask what is stored between sessions, where it lives, who can delete it, and whether background “dreaming” runs on your data.
  • Wrong memory is more dangerous than empty memory. Anthropic’s Pokémon example shows a bad note can drive an agent into a loop. In enterprises that means auditing persistent memory files, not only prompt logs.
  • Dreaming costs compute. Offline consolidation is not free. Budget idle time, rate limits, and which folders get dreaming — especially if the agent can see customer data or an internal wiki.
  • Do not merge three “memory” stories. ChatGPT dreaming, Meta Active Memory, and Anthropic’s agent dreaming are different products/explanations. Keep them apart in procurement and risk reviews.
  • Governance before broad rollout. If you already run Claude Code, Managed Agents, or custom harnesses with persistent folders: define what may live in MEMORY.md-like layers, what must be verified against source, and what must never be auto-injected.

Sources and media

Illustration: AI-generated with Grok Bot / hogby.ai

Mer om AI

Oppskriften. Strategien. Hver dag.

AI-nyheter for ledere. Kuratert av en CIO som bygger det selv. Daglig i innboksen.