Anthropic published a long engineering post on November 6, by Lance Martin, taking apart how Claude keeps knowledge between sessions. It describes a four-layer stack — raw transcripts, notes, linked representations such as wikis, and a small block of injected context loaded into every session — and a process Anthropic calls dreaming: offline time in which Claude rereads its own logs, corrects notes that turned out to be wrong, builds linked wiki pages, and decides which few notes get loaded next time. Its design rule, in Anthropic's words, is to tell Claude what memory is for, but not how to use it.

The numbers are Anthropic's own. On the CL-Bench 1.0 database task, memory lifted Claude Opus 5.5 from 0.655 to 0.838. In a Pokemon Red comparison run — one run per model — Sonnet 3.7 kept 413 memory files and won one badge, while Fable 5 kept nine files and won four. Fed about 350,000 internal files from Stainless, the company Anthropic acquired in May, agents answering 25 questions from a self-built wiki alone rose from 13% to 48% of graded criteria as dreaming scaled from 2 to 128 sessions; wiki plus search scored 61% against 53% for search alone, saving about 20,000 tokens per question.

The post is research, not a product launch: it describes setups on Managed Agents and Claude Code's auto memory, and does not announce dreaming as a general feature. Anthropic is open about the limits — dreaming costs compute, checking every memory continuously is impractical, and a wrong note injected into every session can propagate the same error repeatedly. One practical detail for Claude Code users: it loads only the first 200 lines or 25 KB of the MEMORY.md index when a session starts.