01What is Claude AutoDream and how it works

Claude AutoDream is a background memory consolidation engine built into Claude Code. It runs after your session ends, while you are away or asleep, and organizes what Claude observed during your work. This solves the problem of having to re-explain yourself every time you open an AI tool.

AutoDream runs as a forked sub-agent, a separate process from Claude's main thread. That way, memory work never interrupts your active tasks. It uses a three-gate trigger system.

All three conditions must pass before it activates, which prevents over-dreaming or under-dreaming. When it runs, it goes through four phases: it reads your memory.md file, hunts for new information, consolidates it into memory files, and cleans up contradictions.

02Main features revealed in the leak

The leaked source code revealed several features beyond AutoDream. Here is what is included in the internal architecture:

  • AutoDream: a background memory consolidation engine that runs after sessions.
  • Kairos: an always-on background agent that watches your environment, writes daily logs, and flags important observations.
  • AFK mode: lets Claude Code keep working after you close your laptop, finishing tasks and leaving results ready.
  • Three-layer memory architecture: memory.md for short references, project notes for deeper context, and session history for selective search.
  • Undercover mode: strips Anthropic internal references when working on external projects.
  • Anti-distillation system: injects fake tool definitions into API traffic to prevent competitors from training on Claude Code output.

These features are built but not officially live. Anthropic confirmed the leak was a packaging error.

03How to set up and get started

Since AutoDream is not officially launched, you cannot turn it on yet. But you can prepare now. Here are steps to get your workflow ready:

  1. Set up a memory.md file in your Claude Code project. Include key facts, your preferences, and project context.
  2. Structure your daily logs. AutoDream will prioritize your daily log during its hunt phase, so keep it organized.
  3. Create separate memory profiles for each client or project. This mirrors how AutoDream will build distinct memory files.
  4. Build workflows that can run without your presence. Practice giving Claude Code a task and letting it complete it while you step away.
  5. Connect with other users in communities like the Air Profit Boardroom, where you can learn memory system setup and workflow design.
  6. Watch for official updates from Anthropic. When AutoDream ships, you can flip the feature flag and start using it.

These steps help you get the benefits of persistent memory now, even before AutoDream goes live.

04How to use AutoDream when it arrives

When AutoDream becomes active, here is how it will work in practice. You finish a session with Claude Code. AutoDream runs in the background and processes what you did.

The next time you open Claude Code, it already knows your project, your preferences, and your last decisions. You skip the re-explaining step.

For client work, you could have five clients with different voices and goals. AutoDream builds a separate memory profile for each. When you switch to a client, Claude Code already knows their briefs, feedback, and preferences.

You do not need to brief it again.

AutoDream also connects to kairos. Kairos watches your environment and writes daily logs. AutoDream processes what kairos sees.

Together, they form a feedback loop: kairos observes, AutoDream consolidates, and next session Claude Code knows more about you. Add AFK mode, and you can give Claude Code a task, close your laptop, and come back to completed results.

05Results and comparisons from the episode

The episode did not show live test results because AutoDream is not yet available. But the leak revealed the internal roadmap. The memory architecture is more advanced than what current AI tools use.

Most tools dump everything into the context window at once. Claude Code's system is selective: it knows what to load, what to archive, and where to find information when needed.

The episode compared this to current tools like OpenAI, which still struggle with context and require you to re-explain yourself. With AutoDream, the AI becomes proactive. It gets better over time instead of staying the same.

The episode also mentioned internal model names: Capybara is the potential code name for Claude 5, and Numbap is an unreleased model in pre-launch testing. The latest Capybara build will have better benchmarks, but nothing is confirmed. The memory and agent systems are further ahead than model accuracy right now.

06Tips, limits, and mistakes to avoid

Here are important tips and limits to keep in mind:

  • Do not expect AutoDream to be polished at launch. The episode says it likely will not be fully polished even when it comes out.
  • Do not rely on AutoDream yet. It is not officially active. Anthropic confirmed the leak was a packaging error, and no customer data was exposed.
  • Start building your memory.md now. Structure it so AutoDream has something worth consolidating later.
  • Do not ignore kairos and AFK mode. They work together with AutoDream, so prepare for all three.
  • Avoid dumping all context at once. Use selective memory like the three-layer architecture recommends.
  • Watch for official rollout. Features are built and in the production code base, waiting for a feature flag to flip.

By preparing now, you avoid the mistake of waiting and then scrambling to catch up.

07Who it is for and why it matters

AutoDream matters for anyone who uses AI tools for business. That includes agency owners, solo operators, creators, and e-commerce store owners. If you spend time every session re-explaining who you are and what you are working on, AutoDream saves that time and energy.

The episode argues that this changes the relationship between you and AI. Today, AI is reactive: you drive it, prompt it, and remind it. With AutoDream, Claude Code becomes something that knows your business and your clients.

It gets better over time and runs while you are away. For a solo operator, this could turn a tool into a system that works for you.

Businesses that build these workflows early, before AutoDream becomes mainstream, are likely to compound and grow. Every session the AI knows more, and the gap between them and others widens. The leak gives you a window into what is coming.

08How the leak happened and what to expect next

On March 31st, a developer named XiaoFanShou spotted a source map file on the npm registry. It was a 59.8 megabyte source map accidentally bundled into Claude Code version 2.1 AA. The file was meant for internal debugging only.

Within hours, 512,000 lines of TypeScript were mirrored across GitHub. Thousands of people saw Anthropic's internal code for Claude Code, including AutoDream.

Anthropic confirmed the leak was a packaging error and that no customer data was exposed. The code is real, and the features are built. The leak also exposed internal model names like Capybara and Numbap, plus the undercover mode and anti-distillation system.

What happens next is uncertain. Anthropic is still iterating on the models. The memory and agent systems are further ahead than model accuracy.

AutoDream is in the production code base, waiting for a feature flag to flip. When it does, you can use it right away if you have prepared.