Dream - Memory Consolidation for Claude Code

Your AI agent dreams like you do. Consolidates memory while you sleep.

Anthropic is building an auto-dream feature into Claude Code (currently unreleased). This skill replicates that functionality today - no feature flags, no waiting for rollout. Drop it in and your agent's memory stays clean, current, and contradiction-free.

What It Does

When you use Claude Code across many sessions, auto-memory accumulates noise: stale facts, contradictions, relative dates that lose meaning. Dream fixes this by running a 4-phase consolidation pass over your memory files - the same way your brain consolidates memories during sleep.

Phase 1 - Orient: Reads your current memory directory to understand what exists.

Phase 2 - Gather Signal: Scans recent session transcripts (JSONL files) for user corrections, preference changes, important decisions, and recurring patterns. Uses targeted grep, not full reads.

Phase 3 - Consolidate: Merges new findings into existing memory. Converts relative dates to absolute. Resolves contradictions. Removes references to nonexistent files. No duplicates.

Phase 4 - Prune & Index: Rebuilds MEMORY.md as a lean index under 200 lines. Removes stale pointers. Demotes verbose entries to topic files.

Auto-Trigger

Includes a native Claude Code Stop hook that checks every time you exit a session:

• Has 24+ hours passed since the last dream?
• If yes, flags the next session to run /dream automatically.

Zero overhead when conditions aren't met (~10ms check on exit).

Memory System Auto-Detection

On first install, the skill detects which memory system you're using:

• Native Claude Code (~/.claude/projects/*/memory/) - default
• OpenClaw-style (./memory/ with daily logs)
• Project-root (./MEMORY.md in project root)

If nothing is detected, defaults to native Claude Code memory.

Quick Start

Option 1: Clone into skills directory

git clone https://github.com/grandamenium/dream-skill.git ~/.claude/skills/dream

Option 2: Run the installer

git clone https://github.com/grandamenium/dream-skill.git /tmp/dream-skill
bash /tmp/dream-skill/install.sh --auto

Option 3: Manual install

1. Copy SKILL.md and should-dream.sh to ~/.claude/skills/dream/
2. Run chmod +x ~/.claude/skills/dream/should-dream.sh
3. Start a Claude Code session and say /dream to run it

What's Included

File	Purpose
SKILL.md	The skill prompt - 4-phase consolidation instructions with onboarding
should-dream.sh	Condition checker (24hr timer)
dream-hook.sh	Stop hook that flags next session for dreaming
install.sh	One-command installer with --auto flag for hook setup
test-dream.sh	Creates test fixtures and verifies consolidation results

Usage

Manual

/dream

Automatic (after install --auto)

Just use Claude Code normally. The Stop hook checks on every session exit. When 24 hours have passed, your next session automatically runs a dream consolidation in the background.

Requirements

• Claude Code v2.1.59+ (auto-memory support)
• No additional dependencies

How It Compares to Anthropic's Auto-Dream

Feature	Anthropic (unreleased)	This Skill
4-phase consolidation	Yes	Yes
Session transcript scanning	Yes	Yes
Contradiction resolution	Yes	Yes
Date normalization	Yes	Yes
Auto-trigger	Built into binary	Stop hook + flag file
Memory system detection	Native only	Native + OpenClaw + custom
Available now	Behind feature flag	Yes

License

MIT