How to enable Claude Code dreaming for self improvement

TL;DR

Claude Code can dream by using a nightly dreaming routine to distill sessions, reconcile memories, and fix stale or duplicated notes. The technique described sets up a memory, adds a foresight dream skill, and generates a memory repair report with proposed changes, then applies safe edits while you review. This can dramatically improve task completion by learning from previous days.

Transcript

Andrei Kapathy might be the brightest mind in AI. And 9 months ago, in this interview with Dwarkesh, he revealed what he thought was the biggest current flaw with AI and Claude setups. Roll the clip. >> I feel like when I'm awake, I'm building up a context window of stuff that's happening during the day. But I feel like when I go to sleep, somethin... Read More

Key Insights

• Dreaming is a nightly self learning process that looks for patterns and mistakes across recent sessions and memory content.
• Split focus is a problem where agents try to finish tasks while updating memory, diluting attention and reducing efficiency.
• Patterns obscured occurs when multiple agents write to memory without broad context, leading to misaligned or hidden patterns.
• Memories go stale when duplicates disagree or notes are outdated, guiding decisions with outdated information.
• Anthropic recently released an official dreaming feature, but it is enterprise only and consumes API credits.
• The proposed routine creates a memory system, aligns with a dreaming schema, and adds a foresight dream skill to compare sessions against memory.
• The dream report provides a prioritized, quoted list of proposed changes, with safe, tiny fixes applied automatically where appropriate.
• Running dreaming overnight can significantly boost task completion by consolidating learning across days.

Questions & Answers

Q: What is the basic idea behind Claude Code dreaming and why is it needed?

Claude Code dreaming is a nightly self learning process designed to distill the day’s sessions into a coherent, updated memory. It addresses three major weaknesses: split attention during tasks, patterns obscured across sessions, and memories going stale. By automatically reviewing transcripts and memory content, it can propose corrections and consolidate knowledge, enabling the agent to act more in line with user goals even when not prompted.

Q: How does the video describe the three big problems with memory in Claude Code?

The video identifies three core issues with current Claude Code memory. First, split focus, where the agent divides attention between finishing tasks and updating memory, reducing effectiveness. Second, patterns obscured, the memory file is written by multiple agents without a broader context, causing misaligned patterns. Third, memories go stale, with duplicates and outdated notes that mislead future tasks.

Q: What is the official dreaming feature from Anthropic and what are its limitations?

Anthropic’s official dreaming feature looks for patterns and mistakes across recent sessions and transcripts, creating organized and up to date memory content to enable continuous self learning. The limitations include that dreaming is currently enterprise only and it uses API credits, meaning each dreaming run costs resources and is not available to all users.

Q: What is the purpose of the foresight dream skill in Karpathy’s setup?

The foresight dream skill is designed to read the last 24 hours of session transcripts across different sessions, compare them against the current memory, and identify corrections. It then proposes changes in a numbered list, quotes evidence from transcripts, and applies only tiny safe fixes automatically. This helps unify patterns and remove stale or conflicting memories.

Q: How is the dreaming routine scheduled and executed according to the video?

The routine is scheduled to run at 3:00 a.m. every night, using whatever the computer supports for timing. A test run can be done immediately to show the proposed changes. If running overnight with no activity, it writes the proposals to a dream report for morning review. The process can also auto apply safe fixes if preferred.

Q: What benefits did the enterprise customers reportedly experience with dreaming?

Enterprise customers Harvey and Rakuten reportedly saw a sixfold increase in task completion rates after enabling the dreaming feature. This suggests that continuous self learning from nightly dreams substantially improves performance by better aligning actions with recent goals and patterns across sessions.

Q: What is the role of the dream report in Karpathy’s workflow?

The dream report serves as a reviewable, structured summary of proposed memory changes. It provides a numbered list of corrections, a short transcript quote as evidence, and options to approve or reject each change. It enables the user to oversee memory updates, ensuring only safe, minor edits are applied automatically while larger changes are reviewed.

Q: What is the overall goal of introducing a dreaming routine for Claude Code?

The overall goal is to enable Claude Code to learn from all user sessions even when not actively prompted, by distilling experiences into memory and updating behavior to better reflect user goals. This reduces stale information, improves pattern recognition across sessions, and increases task completion efficiency by delivering a more cohesive, evolving intelligence.

Summary & Key Takeaways

• This video explains the concept of dreaming for Claude Code, a nightly process that analyzes recent sessions and memory to produce up to date content, reducing stale information and improving performance. It includes a practical prompt setup and a live demonstration of the dream report in action.

• The presenter contrasts current Claude memory issues with the newly demonstrated dreaming routine, highlighting how patterns can be detected across sessions, duplicates removed, and memories synchronized, leading to better alignment with user goals.

• A step by step walkthrough shows how to implement a free dreaming prompt, schedule nightly runs at 3:00 a.m., and review a dream report that proposes changes with evidence quotes from transcripts.