Claude Code has an unreleased feature called Auto-dream. I found it, I can't use it, and I think that's fine.
Claude Code is building a memory consolidation system modeled on human sleep. It's not available yet. Here's what it does and why I'm not waiting.
Someone on Reddit noticed a toggle in Claude Code's /memory menu labeled "Auto-dream: off." It can't be turned on. Anthropic built the bedroom but hasn't handed out the pajamas yet.
Here's what it actually does when it eventually works.
Auto memory -- the feature that lets Claude Code take notes about your project during sessions -- has a slow decay problem. After 20 or 30 sessions, those notes become a mess. Contradictory entries. Relative timestamps like "yesterday we switched to Fastify" that lose meaning over time. Stale debugging notes for files that no longer exist. The notebook that was supposed to help Claude remember starts to confuse it instead.
Auto-dream is the cleanup crew. It runs a background agent periodically -- minimum every 24 hours, after at least 5 sessions have accumulated -- and consolidates everything. Merges overlapping entries. Converts "yesterday" to actual dates. Deletes facts that got contradicted. Keeps the MEMORY.md index under 200 lines because that's the cutoff for what loads at session start.
The analogy Anthropic seems to be reaching for is human sleep. Daytime: your brain absorbs raw input. REM: it replays the day, strengthens what matters, discards what doesn't. Auto memory is the daytime brain. Auto-dream is the sleep cycle.
It's controlled by a server-side feature flag. Not a settings toggle you can flip. Anthropic rolls it out on their end.
I checked. My instance has the toggle. It's greyed out.
Why I'm not waiting for it
I'm starting a new project this week -- a daily AI morning digest built in a separate repo from this site. New repo, zero session history. Auto-dream needs accumulated sessions before it has anything to consolidate. CLAUDE.md and AGENTS.md do the job better at the start of a project anyway.
When Auto-dream goes GA and I've got 20 sessions of digest-building context piled up, I'll enable it. That's the right moment. Using it now would be enabling a janitor before the house has any furniture.
What actually matters here
The X post that surfaced this got nearly 57,000 views in under 24 hours. Someone noticed an unreleased toggle, dug into the binary, found the feature flag name (tengu_onyx_plover, for the curious), reverse-engineered the consolidation logic, and wrote it up. The whole thing happened in a day.
That's the pace right now. Features ship before they're announced. The people paying attention to the source code know about things before the changelog does.
I want a system that catches exactly this kind of signal every morning without me having to spend an hour on Reddit finding it. That's what I'm building next.
Session cost: writing only. No Claude Code tokens. Just a thing I noticed and wanted to document before I forgot why I decided not to use it.