Anthropic details Claude agent memory and offline dreaming

Anthropic wrote on X on 6 November 2026 that memory is what lets Claude learn from experience, and that a new engineering post explains how “dreaming” helps the model organize what it has learned.

“Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time.”

— Anthropic (@AnthropicAI), 6 November 2026

The linked research page, The components of agent memory, is Anthropic’s own account of how it structures long-running agent context. The company describes layers that include searchable session transcripts, memory notes, structured representations, and a small injected-context set loaded at session start. Dreaming is framed as offline processing that validates notes against transcripts, corrects errors, and promotes important items into injected context.

What this is not

The post is an engineering explanation, not a new model release and not a claim of human-like sleep. Anthropic is describing software memory machinery for Claude agents. Readers should treat performance claims and design choices as the lab’s self-description until independent teams reproduce the setup.

For the public record, start with the 6 November X announcement and the engineering writeup.

This report is not legal, financial, or health advice. See the disclaimer.