Anthropic explains memory in Claude agents
Anthropic describes how Claude agents can retain and organize useful information across sessions.
- The design combines logs, notes, references and context reused across sessions.
- Anthropic reports background review cycles that consolidate information and correct errors.
- In a Pokémon Red test, Fable 5 used 10 files and progressed further than Sonnet 3.7, which created 413.
Context and analysis by Noosigma AI
Editorial summary and context based on the cited source. Claims and results attributed to the organization are reports from that source; they do not constitute independent validation by Noosigma AI.
The article separates memory into session logs, notes with salient observations, more structured references and context that can be loaded at the start of a future task. A background review step, called “dreaming” in the post, reorganizes this material and can correct errors or produce new notes.
Why it matters
On CL-Bench tasks, Anthropic reports that some models recorded wrong answers, tested possible causes and saved verified rules for later questions. In Pokémon Red, Claude Fable 5 progressed further with 10 memory files and earned four badges; Sonnet 3.7 created 413 files and earned one.
The examples and evaluations were reported by Anthropic and do not establish how other agents would perform on different tasks. Source: Anthropic, November 6, 2026.