Anthropic didn’t announce this one. There was no blog post, no TechCrunch headline, no product launch email. AutoDream was discovered by developers poking around Claude Code’s internals, reading system prompts they weren’t supposed to see. The system prompt says it plainly: “You are performing a dream — a reflective pass over your memory files.”
That’s either the most poetic thing a software tool has ever said, or a sign that someone at Anthropic is having too much fun naming internal features. Either way, the feature is real, it’s rolling out now, and it fixes something that’s been quietly broken about Claude Code’s memory for a while.
The Problem with AutoMemory#
AutoMemory — the predecessor — let Claude Code persist notes across sessions. The idea was right: long-running projects benefit from context that outlasts individual conversations. You don’t want to re-explain your architecture every time you open a new session.
The execution had problems.
Memory files decayed. Notes written three weeks ago referenced “yesterday’s refactor” with no absolute date. Conflicting decisions piled up without resolution — Claude would have two notes that said opposite things about the same module and no way to know which was current. And memory files grew unbounded, consuming token budget that should have gone to actual code.
The result was a memory system that technically worked but required manual curation to stay useful. Which meant most people didn’t curate it, which meant the memory got progressively less reliable over time.
What AutoDream Does#
AutoDream runs as a background sub-agent between sessions — never during active work. It executes a four-phase process:
- Orientation — reads the current memory files and assesses their state
- Gather Signal — identifies what’s stale, conflicting, or redundant
- Consolidation — merges fragmented or related notes into coherent entries
- Prune and Index — removes outdated entries and restructures the file for retrieval efficiency
The concrete example that keeps coming up: relative temporal references. AutoDream converts “Yesterday we decided to use Redis” into “On 2026-03-15, we decided to use Redis.” Small change. Enormous difference in how useful that note is three months from now.
It also resolves conflicts. When two memory entries contradict each other, AutoDream determines which is more recent and more likely to be current, then consolidates them into a single authoritative note. No more Claude confidently citing a decision that was reversed six commits ago.
Why the Sleep Metaphor Is Actually Apt#
The human brain consolidates memory during REM sleep — transferring short-term experiences into long-term storage, pruning irrelevant details, strengthening connections between related concepts. This is well-established neuroscience, and it’s a legitimately good metaphor for what AutoDream does.
Claude Code’s equivalent: you work on a project for an hour, generate a bunch of session notes, and AutoDream runs afterward to decide what’s worth keeping, what’s stale, and how everything fits together. The agent that starts your next session is working from a cleaner, more accurate picture of your project than the one that ended the last session.
The “between sessions” timing is deliberate. You don’t want a background process competing for tokens while you’re actively working. And the consolidation benefits from having a complete session to work with rather than interrupting one mid-stream.
Rollout Status (Honest Assessment)#
This is still a staged rollout with some rough edges.
The feature flag is tengu_onyx_plover on the server side — not a user-configurable toggle that’s universally available yet. The minimum interval between dream runs is 24 hours. It requires at least 5 accumulated sessions before it triggers. Some users have reported that autoDreamEnabled shows true in their settings and the toggle appears in /memory, but running /dream manually returns “Unknown skill” — the server-side flag hasn’t flipped for them yet.
Anthropic has said public rollout is coming within the week. Given their track record with Claude Code launches, that probably means sometime in the next 72 hours.
To check if you have it: run /memory inside a Claude Code session and look for “Auto-dream: on” in the selector. You can also add "auto_dream": true to ~/.claude/settings.json — this won’t force the feature if your account isn’t in the rollout, but it’ll enable it automatically when you are.
What This Changes#
The practical impact is larger than the feature description suggests.
Claude Code has been positioning itself as a long-running project collaborator, not just a task-by-task tool. That positioning only works if the memory system is reliable. Unreliable memory means you spend the first few minutes of every session correcting Claude’s misconceptions about your codebase. That’s the opposite of delegation.
AutoDream is what makes the memory system trustworthy over time. It’s the maintenance layer that ensures the context Claude starts with is accurate, not just accumulated. The difference between a junior dev who takes good notes and one who takes good notes and reviews them regularly.
This also has implications for context efficiency. Pruned memory files use fewer tokens. On long projects, that compounds — you’re not burning context budget on stale information from two months ago.
The Competitive Gap This Opens#
Cursor uses static .cursorrules files. Useful for project setup, not designed to evolve. GitHub Copilot has minimal persistent memory. Windsurf has project-level rules but no cross-session learning mechanism.
AutoDream is the first AI coding tool that actively manages its own memory over time. That’s a meaningful capability gap, and it compounds — a Claude Code instance that’s been working on your project for three months will be qualitatively more useful than a fresh setup, not just quantitatively. The memory gets better as it gets curated.
Whether that gap stays wide depends on how quickly the other tools ship comparable features. But Anthropic has a structural advantage here: Claude Code’s memory system was designed as a first-class feature from the start, not retrofitted onto a chat interface. AutoDream is a natural extension of that architecture.
The Unannounced Launch Tells You Something#
Anthropic didn’t announce AutoDream because it wasn’t ready for a public launch. The staged rollout, the buggy /dream command, the server-side flag — these are signs of a team shipping something real before it’s polished, which is the right call. Waiting for perfect means shipping late.
But it also reflects something about where Claude Code is in its development cycle: features are landing faster than the communications team can announce them. That’s a good problem to have, and it means the changelog is worth reading more carefully than the press releases.
AutoDream will get a proper announcement soon. For now, check /memory, enable the flag if you have access, and let your agent sleep.
Sources: Claude Code AutoDream: Anthropic’s New Memory Feature — ClaudeFa.st, Does Claude Code Need Sleep? Inside the Unreleased Auto-dream Feature — DEV Community, What Is Claude Code AutoDream? — MindStudio, AutoDream GitHub Issue #38461 — anthropics/claude-code
July 2026 Update: AutoDream Never Got Its Announcement. It Got Split in Two Instead.#
Four months on, the “proper announcement soon” line above didn’t age well — at least not the way it implied. /dream never shipped as a documented Claude Code command, and tengu_onyx_plover never graduated into a public toggle. Search the current Claude Code changelog for “dream” or “AutoDream” and you’ll find nothing. The client-side experiment this article covered appears to have been quietly retired rather than launched.
What happened instead is more interesting than a straight GA: Anthropic productized the same underlying idea one layer up, on the Claude Managed Agents platform, and split it into two separate primitives.
Memory launched in public beta on April 23, 2026 — filesystem-based, cross-session persistence for managed agents, with API access via the managed-agents-2026-04-01 header. Early adopters cited in Anthropic’s own announcement include Netflix, Rakuten, Wisedocs, and Ando; Wisedocs-style document-verification workflows reported a 97% reduction in first-pass errors and a 30% speed increase. As of this update, Memory is still public beta, not GA — Anthropic was still expanding webhook coverage for it as recently as July 22.
Dreaming followed at Code with Claude SF on May 6 — the direct conceptual descendant of what this article called AutoDream: a scheduled, overnight review of past agent sessions and memory stores that merges duplicates, removes outdated entries, and surfaces recurring patterns. This blog covered it the same week (Claude Agents Can Now Dream. Harvey Saw 6× More Tasks Completed.), citing Harvey’s reported 6x completion-rate increase. Ten weeks later, Dreaming remains in research preview — third-party conference recaps single it out as the one primitive from that May announcement (alongside Outcomes and multiagent orchestration) that still hasn’t reached general availability. No numbers beyond the original Harvey figure have surfaced since.
Two things are worth separating here. First: the original AutoDream feature — the one embedded directly in interactive Claude Code sessions, with the REM-sleep framing — never got the “proper announcement” this article predicted. It’s not in the docs, and the closest thing to it, /memory, just opens your CLAUDE.md/MEMORY.md file for manual editing. Second, and more consequential: what Anthropic actually shipped is memory management for the Managed Agents platform — autonomous, API-driven agents running on Anthropic’s infrastructure, not the interactive terminal session you drive by hand. That’s a different product surface than the one this article was written about.
What Claude Code itself has gained since is more modest but real: MEMORY.md is now first-class infrastructure. The changelog shows a steady stream of small, unglamorous fixes — v2.1.181 (June 17) trimmed verbose per-file memory listings, v2.1.186 (June 22) added a reminder to compact the memory index near its size limit, v2.1.203 (July 7) fixed the over-limit warning to measure only loaded content, v2.1.210 (July 14) turned silent truncation on oversized writes into a hard error, and v2.1.214 (July 18) added a modified timestamp to memory file frontmatter. None of it is “your agent dreams” — all of it is the unglamorous work of making a growing memory file behave predictably, which is arguably the more important half of the original problem this article described.
The competitive gap this article predicted has held, though the shape of it moved. Windsurf effectively dissolved as an independent product this spring; Gemini CLI was sunset for consumer tiers on June 18 in favor of Antigravity. Cursor and GitHub Copilot CLI still rely on static or manually-curated memory files with no scheduled consolidation loop — a third-party community layer (“agentmemory”) has emerged specifically to bolt cross-session memory onto Copilot CLI, Cursor, and Gemini CLI, which is itself evidence nobody else has shipped a first-party equivalent. Anthropic remains the only vendor with a self-maintaining memory system in production anywhere in its stack — it’s just running on the Managed Agents side of the house, not the one this article originally covered.