Claude Code's Unreleased 'Auto-Dream' Feature: AI Memory That Consolidates Overnight
Buried in Claude Code's /memory settings panel is a feature that hasn't been publicly announced yet: Auto-dream, currently showing status "off never" — present in the code but not yet enabled for users. Based on Reddit discussions and the claudefa.st reference documentation, Auto-dream works by periodically spawning a background subagent that reorganizes and consolidates your memory files: archiving high-value long-term context, pruning redundant entries, and converting relative date references into absolute timestamps. That last detail matters more than it sounds — memory files that say "yesterday we decided to use Redis" become meaningless within a week; files that say "On 2026-03-28 we decided to use Redis" stay useful across month-long projects.
The concept draws direct inspiration from Cursor's Continual Learning plugin, which uses an agents-memory-updater subagent to extract context from conversation history into AGENTS.md files. Anthropic appears to be building an equivalent but in the form of autonomous overnight consolidation rather than explicit user-triggered saves. The "off never" flag indicates this is in active internal testing, not just an architectural sketch — most developers familiar with the codebase read that status pattern as indicating an imminent public preview. For anyone running long Claude Code sessions where context management is already a pain point, this is the feature to watch.