TechInsights7 min de lecture

Claude Code Dream & Auto Dream: Automatic Memory Consolidation

Thomas Gros•
Illustration du système Auto Dream dans Claude Code consolidant la mémoire entre les sessions

What is Auto Dream in Claude Code?

TL;DR: Auto Dream is a Claude Code sub-agent that automatically consolidates memory files every 24h (after 5+ sessions). It deduplicates entries, removes stale notes, converts relative dates to absolute, and keeps MEMORY.md under 200 lines. Feature rolling out gradually (March 2026).

Auto Dream is an automatic memory consolidation mechanism in Claude Code. It works as a background sub-agent that cleans, organizes, and updates the memory files accumulated by Auto Memory across sessions.

The name draws from neuroscience: during REM sleep, the brain replays the day's events, strengthens useful connections, and discards what doesn't matter. Auto Dream does the same thing with Claude Code's notes.

The problem: memory rot

Auto Memory is a powerful feature. Claude takes its own notes while working: build commands, architecture decisions, style preferences, resolved bugs. These notes live in ~/.claude/projects//memory/ and are reloaded every session.

The problem appears after 20, 50, or 100 sessions. Notes accumulate without maintenance:

  • Stale relative dates: "Yesterday we decided to use Redis" loses all meaning three weeks later
  • Contradictory entries: one note says "API uses Express," another says "Fastify migration complete"
  • Duplicates: three different sessions noted the same build command
  • Obsolete notes: debug solutions for files that no longer exist
  • Overloaded MEMORY.md: the index exceeds 200 lines, and anything beyond that isn't loaded at startup

The result: Claude starts hallucinating based on outdated information, or loses important instructions buried in noise.

How does Auto Dream work?

Auto Dream runs in four phases, inspired by the brain's sleep memory consolidation process.

Phase 1: Orientation

Claude scans the memory directory and reads MEMORY.md to understand the current state. It builds a map of what exists.

Phase 2: Signal gathering

The sub-agent searches through recent session transcripts (JSONL files) to identify high-value patterns: user corrections, explicit saves, recurring themes, important decisions. It doesn't read everything exhaustively; it searches for specific patterns.

Phase 3: Consolidation

This is the core work:

  • Date conversion: "yesterday" becomes "2026-03-24." Absolute dates remain readable over time.
  • Contradiction removal: if you migrated from Express to Fastify, the "API uses Express" entry is removed
  • Duplicate merging: three notes about the same build command become one clean entry
  • Stale note cleanup: debug solutions for deleted files are removed

Phase 4: Pruning and indexing

MEMORY.md is updated: stale pointers removed, new files referenced, contradictions resolved. The index is kept under 200 lines (the startup loading threshold).

Trigger conditions

Auto Dream doesn't run every session. Two conditions must both be met:

ConditionThresholdReason
Time elapsed24 hours minimum since last consolidationAvoid unnecessary runs
Sessions accumulated5 sessions minimum since last consolidationEnsure enough material to consolidate

This dual gate prevents runs on inactive projects while ensuring regular cleanup on active ones. One observed case consolidated 913 sessions' worth of memory in about 8-9 minutes.

Safety guarantees

Auto Dream operates under strict constraints:

  • Read-only on code: the sub-agent can only modify memory files, never source code
  • Lock file: prevents concurrent consolidation on the same project
  • Background execution: never blocks the active session

Manual triggering

You can force consolidation at any time by asking in conversation:

dream
consolidate my memory files
auto dream

Claude will run the consolidation process immediately, without waiting for the 24h and 5-session thresholds.

Claude Code's four memory systems

Auto Dream fits within an ecosystem of four complementary memory mechanisms:

SystemRoleTriggerWritten by
CLAUDE.mdPersistent project instructionsManualYou
Auto MemorySession learning captureContinuousClaude
Session MemoryConversation continuity across sessionsAutomaticClaude
Auto DreamConsolidation and maintenanceEvery 24h + 5 sessionsClaude

CLAUDE.md sets the rules. Auto Memory takes notes. Session Memory maintains the conversation thread. Auto Dream does the housekeeping.

Concrete impact

Before consolidation: a memory directory with a 280+ line MEMORY.md, contradictory entries, references to deleted files, duplicates spread across multiple files.

After consolidation: MEMORY.md drops to ~142 lines (under the 200-line startup threshold), contradictions resolved, all dates absolute, redundant entries merged into single authoritative files.

The signal-to-noise ratio improves, and Claude works more effectively on long-running projects.

Availability

Auto Dream is rolling out gradually (March 2026). The feature is controlled by a server-side feature flag. It appears in the /memory menu but can't yet be manually enabled by all users.

Configuration parameters reveal the approach:

{
  "minHours": 24,
  "minSessions": 5,
  "enabled": false
}

The enabled parameter is managed server-side. Local settings cannot override it.

Theoretical foundation

Auto Dream's design builds on the research paper "Sleep-time Compute: Beyond Inference Scaling at Test-time" (April 2025). This research demonstrated that preprocessing context during idle periods can reduce inference costs by a factor of 5.

By consolidating memory between sessions rather than during active work, Auto Dream applies this principle: the agent's "sleep" time is useful compute.

Learn more

Memory management is just one aspect of Claude Code. Our one-day Claude Code training covers CLAUDE.md files, Plan Mode, subagents, hooks, and custom skill creation.

Frequently Asked Questions

Auto Memory captures learnings during work sessions. Auto Dream consolidates and cleans those notes between sessions. Auto Memory writes; Auto Dream maintains.
Auto Dream only removes contradicted entries (replaced by more recent information), duplicates, and obsolete notes referencing deleted files. Memory files remain editable Markdown. You can review and correct anytime via /memory.
Currently, Auto Dream runs during active Claude Code sessions. Anthropic already has cloud scheduled task infrastructure that could enable offline execution in the future.
The feature is in gradual rollout and isn't enabled for all users yet. Auto Memory can be disabled via /memory or by adding "autoMemoryEnabled": false to your settings.
Auto Dream runs in the background as a sub-agent. Consumption depends on the volume of memory to consolidate. The documented case of 913 sessions took 8-9 minutes of processing.
Memory files are Markdown in ~/.claude/projects//memory/. You can inspect changes with git diff if you version that directory, or simply read the files via /memory.

Tags

#claude-code#agentic-coding#productivite#ia#memoire#automatisation