By AutoDream is a background memory consolidation feature for Claude Code that prunes, merges, and reorganizes your project's memory files between sessions. I've been watching it closely since it surfaced in the /memory menu this week, and the short version is: if you've been using AutoMemory and noticed your agent getting dumber over time, this is the fix.
What AutoDream actually does
Claude Code has had AutoMemory since v2.1.59 in late February 2026. AutoMemory lets Claude take notes as it works: build commands, architecture decisions, debugging patterns, your coding preferences. The problem is that AutoMemory only writes. It never cleans up. After 20-plus sessions, your memory files are full of contradictions, stale references to deleted files, and five different entries that all say "use TypeScript strict mode."
AutoDream is the cleanup crew. It runs a background sub-agent that consolidates those accumulated notes, and the metaphor Anthropic chose is deliberate: it works like REM sleep. AutoMemory is the daytime brain, absorbing information. AutoDream is the sleep cycle that replays, organizes, and discards.
The process follows four phases, as documented by claudefa.st:
- Orientation -- Claude reads the current memory directory and builds a map of what exists
- Gather signal -- It searches session transcripts for high-value patterns: user corrections, explicit "remember this" moments, recurring themes, and architecture decisions. It uses targeted searches rather than reading every transcript end-to-end
- Consolidation -- The core phase. It converts relative dates to absolute ones ("yesterday" becomes "on 2026-03-15"), deletes contradicted facts, removes stale entries, and merges overlapping notes
- Prune and index -- Updates the MEMORY.md index file, which is capped at 200 lines since that's the startup loading cutoff
How to access it (and why you probably can't yet)
Here's the catch: AutoDream is controlled by a server-side feature flag. According to a dev.to investigation by Akari Iku, the flag's codename is "tengu_onyx_plover" and Anthropic manages the rollout on their end. You can see the toggle in your /memory menu, but for most users it reads "Auto-dream: off" and you can't flip it on.
The default parameters tell you a lot about how Anthropic thinks about this:
- minHours: 24 -- it runs at most once per day
- minSessions: 5 -- it won't trigger until you've accumulated five sessions worth of new notes
- enabled: false -- server-side, not something you toggle locally
You can trigger a one-off consolidation manually with the /dream command inside Claude Code. That's available now regardless of the feature flag, and it's worth trying if your CLAUDE.md has gotten unwieldy.
Why this matters for long-running projects
The core tension here is real and well-understood: language models have finite context windows, and every token your memory files consume at session start is a token unavailable for actual work. MindStudio's breakdown puts it well -- your agent's effective memory, the portion that's actually useful, shrinks as a fraction of the whole if nobody curates it.
I've experienced this firsthand. On projects where I've run 30-plus Claude Code sessions, the CLAUDE.md file becomes a wall of partially redundant notes. Claude starts contradicting itself because its own memory contains conflicting entries from different stages of the project. It's the AI equivalent of a notebook where you never tear out old pages.
AutoDream addresses this structurally rather than asking you to manually prune your memory files every few days. The "gather signal" phase is particularly smart: instead of re-reading every session transcript (which would burn through tokens fast), it does targeted searches for corrections, explicit saves, and recurring patterns.
How it compares to manual memory management
Before AutoDream, the best practice was to periodically open your CLAUDE.md and memory files, delete outdated entries, and reorganize by hand. Some developers wrote custom scripts to deduplicate their memory directories. A few turned off AutoMemory entirely because the noise got worse than the signal, as the dev.to article's author noted.
Cursor and Windsurf handle project context differently. Cursor uses .cursorrules files that are entirely manual, with no automatic note-taking and no consolidation. Windsurf's Cascade has session memory but it resets between conversations rather than persisting long-term. Neither has attempted the write-then-consolidate loop that AutoMemory plus AutoDream creates.
That loop, automatic note-taking followed by automatic curation, is genuinely novel in the coding agent space. It's the closest any tool has come to maintaining useful project context across dozens of sessions without human intervention.
The research angle
The theoretical basis isn't hand-waving. AutoDream's design aligns with a paper from April 2025 called "Sleep-time Compute: Beyond Inference Scaling at Test-time" by Kevin Lin, Charlie Snell, and collaborators at Letta and UC Berkeley. The paper argues that LLMs can benefit from compute spent between queries, not just during them, analogous to how sleep consolidation improves human memory retrieval.
Anthropic hasn't publicly cited this paper in connection with AutoDream, but the timing and design philosophy line up closely.
Pricing and cost considerations
Anthropic hasn't published separate pricing for AutoDream. Since it runs as a sub-agent using Claude, it will consume tokens from your existing Claude Code usage. The once-per-day, five-session-minimum defaults suggest Anthropic is being conservative about token burn. The targeted search approach in phase two (grep-style lookups rather than full transcript reads) also keeps costs down.
Claude Code itself requires a Max plan ($100/month for individuals, $200/month for teams) or API access. AutoDream doesn't appear to add any additional cost beyond the tokens the consolidation agent consumes during its runs.
Buy, skip, or wait?
Wait, but actively. AutoDream solves a real problem that anyone running AutoMemory on long-lived projects has hit. The four-phase consolidation process is well-designed, and the sleep metaphor isn't just marketing -- it maps to a genuine computational insight about between-query processing.
But it's not fully shipped yet. The feature flag means most users can't turn it on, and there's no timeline for general availability. In the meantime, use /dream manually after heavy project sessions. If your CLAUDE.md is already bloated, run it today.
For teams evaluating Claude Code against Cursor or Windsurf for long-term projects, AutoDream tilts the comparison meaningfully. No competing tool has anything like automatic memory curation. If your projects span months and dozens of sessions, this is the kind of feature that compounds in value.
Marcus Webb covers AI products for The Daily Vibe.
