Anthropic explains how agent memory works, including offline dreaming

Anthropic published an engineering post titled The components of agent memory on 6 November 2026, describing how Claude retains knowledge between sessions through notes, retrieval, wikis and an offline process it calls dreaming. As Anthropic scaled dreaming from two sessions to 128, scores on 25 questions grew from 13% to 48% of graded criteria. [1]

What does Anthropic mean by agent memory?

Anthropic published The components of agent memory on 6 November 2026, covering notes, wikis, retrieval and dreaming. [1] By agent memory, Anthropic means knowledge Claude gains from experience and keeps between sessions as context it can read later, not changes to model weights, in-context learning inside one conversation, or compaction summaries that replace older turns.

Anthropic says agents already capture much of their work in logs or transcripts. Researchers at Letta and UC Berkeley proposed sleep-time compute as a way for models to use idle time between interactions to reorganize context; Anthropic says the same idea can help agents fix mistakes or produce new notes. [1]

What is dreaming, and what did the document test show?

Dreaming, in Anthropic’s telling, is offline reflection on those logs to correct errors, write notes, build new memory representations, or decide what to load into a later session. [1] One representation is a wiki: pages Claude writes and maintains about a subject, with links that help it find related notes.

Anthropic explored wikis on a corpus of about 350,000 internal documents, roughly 420 million tokens, from Stainless, a software company Anthropic says it acquired in May. [1] As Anthropic scaled dreaming from two sessions to 128, scores on 25 questions grew from 13% to 48% of graded criteria.

Anthropic says Claude’s best score was 48% when it used only the wiki to answer those questions. Using search and the wiki raised scores from 54% to 61% as the wiki grew, compared with 53% with the search tool alone, and saved about 20,000 tokens per question. [1]

How did memory change a Pokémon Red run?

Anthropic also evaluated what Claude judged worth saving while playing Pokémon Red, comparing one run each of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 steps. [1] In a Pokémon Red test, Claude Sonnet 3.7 earned one badge while Claude Fable 5 earned four badges.

Sonnet 3.7 wrote 413 memory files across 225 folders and, Anthropic says, spent growing time combing through stale notes because the game state changes constantly. [1] Fable 5 wrote nine files in two folders, plus a top-level file called global.md.

Our read

Opinion: The useful claim here is not that Claude suddenly has a hippocampus, but that Anthropic is treating memory as an engineering stack—notes, retrieval, wikis and offline cleanup—and publishing evaluations that show both gains and failure modes.

What we could not verify

• Independent replication of the Stainless wiki scores and the CL-Bench learning-loop comparison.
• Whether the Pokémon Red result holds beyond one run per model at about 10,000 steps.
• How much of the dreaming recipe Anthropic uses in customer-facing Claude products today.