Claude 'dreams' to tidy its memory: Anthropic says wiki score rose from 13% to 48% as dreaming sessions scaled to 128

On Nov. 6 Anthropic published "The components of agent memory" on its engineering blog, introducing "dreaming," a process in which Claude reviews past session transcripts between sessions to organize memory. The official post drew about 2,800 likes. The name comes from how people consolidate experience during sleep.

According to the blog, dreaming checks memory notes against the original transcripts to fix errors, builds linked wiki pages of what Claude knows, and promotes key notes into the context loaded at the start of each session. In a Pokemon test, a wrong note about a building's location led Claude off a ledge; after dreaming corrected it, Claude reached the next city. On a corpus of about 350,000 files (roughly 420 million tokens) with 25 questions, wiki-only scores rose from 13% to 48% as dreaming scaled from 2 to 128 sessions. Wiki plus search reached 61% versus 53% for search alone and saved about 20,000 tokens per question. On a contract-terms task, coverage of graded facts rose from 72% (41 of 57) to 100% after Anthropic changed how agents wrote the wiki.

The blog lists limits too. On one single-document task the wiki showed no clear gain, and on another it grew to 1.77 times the size of the source files. It does not say who can use dreaming or when, and points to Managed Agents documentation and getting-started guides. Replies ranged from "much needed memory update" to the joke "Claude isn't the only one dreaming."

The same day, ClaudeDevs said it made it easier to get a repo running in Claude Code cloud sessions on Pro, Max and Ultra (about 3,700 likes). Claude figures out the dependencies, network access and API keys it needs and saves the environment; API keys are stored securely and Claude never sees them.