Descriptions:
Anthropic has quietly rolled out an experimental feature for Claude Code called AutoDream, which runs a background sub-agent that periodically consolidates and prunes Claude’s memory files — analogous, the video notes, to how human brains consolidate memories during sleep. In this walkthrough, Nate Herk activates and demonstrates the feature across multiple projects, including one with 285 sessions of accumulated history, observing the process in real time via the task view and status line.
AutoDream builds on Claude Code’s existing auto-memory system, which stores project context in markdown files injected at the start of each session. The difference is that AutoDream actively manages those files over time: it reads recent session data, runs a consolidation pass through a dedicated sub-agent, and outputs cleaner, more compact memory files by removing redundancy, trimming outdated entries, and reorganizing what remains. In Herk’s demo, a run across 13 sessions took roughly 10 minutes and improved five separate memory files.
Practically, users can enable AutoDream from the `/memory` menu in Claude Code, toggle it per project, and manually trigger it with `/dream`. Herk walks through the concrete benefits: less repetition across sessions, reduced context bloat, and sharper recall because the memory index stays focused rather than growing unbounded. He also speculates on the internal structure of the dream prompt, suggesting it functions as a reflective synthesis pass. The feature is still in gradual rollout and exhibits some experimental behavior, but early results point toward meaningfully improved long-term session continuity.
📺 Source: Nate Herk | AI Automation · Published March 24, 2026
🏷️ Format: Tutorial Demo
