Sunday, November 8, 2026 · Claude & Anthropic
Claude · Research · happened Nov 6, 2026
Anthropic details how Claude 'dreams' between sessions to tidy agent memory, lifting a wiki test from 13% to 48%
In a Nov. 6 research post, Anthropic describes dreaming, where Claude reviews past transcripts between sessions to fix wrong notes and write new ones. In its own tests, a wiki Claude built over 128 dreaming sessions from about 350,000 internal Stainless files raised scores on 25 questions from 13% to 48%, and in Pokémon Red, Claude Fable 5 kept about 10 memory files and earned four badges, where Sonnet 3.7 wrote 413 files and earned one. Anthropic notes the wiki helped less when a plain text search could find the answer in one document.
Why it matters
Memory that corrects itself is one of the main limits on long-running agents, and these are early, mostly single-run results from Anthropic itself.
Try it
Dreaming is a research preview in Claude Managed Agents on the Claude Platform; developers request access from Anthropic.