🌙 OpenClaw Autodream

Automatic memory consolidation for OpenClaw agents — like REM sleep for AI memory.

Autodream analyzes your agent's daily memory files, removes duplicates, prunes stale entries, normalizes relative dates, and builds a clean, organized MEMORY.md index. Inspired by Claude Code's Autodream feature.

The Problem

OpenClaw agents accumulate daily memory files (memory/YYYY-MM-DD.md) over time. Without consolidation:

• Duplicates pile up — same information recorded across multiple days
• Stale entries persist — completed tasks, resolved bugs, outdated workarounds
• Relative dates decay — "yesterday" and "last week" become meaningless
• No organization — insights scattered across dozens of files
• No long-term memory — agents read recent files but miss older patterns

The Solution

Autodream runs a 4-phase consolidation pipeline (no LLM required):

1. Orientation — Maps current memory state, identifies files to process
2. Gather Signal — Extracts and classifies entries from daily files
3. Consolidation — Deduplicates, prunes stale entries, normalizes dates
4. Prune & Index — Enforces size limits, writes organized MEMORY.md

Installation

npm install -g openclaw-autodream

Or clone and link:

git clone https://github.com/rimaslogic/openclaw-autodream.git
cd openclaw-autodream
npm link

Usage

# Check if consolidation is needed
autodream ~/.openclaw/workspace --stats

# Preview what would change (dry run)
autodream ~/.openclaw/workspace --dry-run --verbose

# Run consolidation
autodream ~/.openclaw/workspace --verbose

# Force full reconsolidation (ignore trigger conditions)
autodream ~/.openclaw/workspace --force --verbose

# Custom limits
autodream ~/.openclaw/workspace --max-lines 150 --lookback 60

Options

| Flag | Description | |------|-------------| | --dry-run | Preview changes without writing files | | --force | Ignore trigger conditions (24h + 5 files) | | --verbose | Detailed output during consolidation | | --stats | Show memory statistics and exit | | --max-lines N | Maximum lines for MEMORY.md (default: 200) | | --lookback N | Days to look back for daily files (default: 30) | | --help | Show help message | | --version | Show version |

How It Works

Phase 1: Orientation

Reads existing MEMORY.md (if any), scans the memory/ directory, and identifies files changed since the last consolidation.

Phase 2: Gather Signal

Parses each daily file and extracts structured entries. Each entry is:

• Classified into categories (People, Projects, Preferences, Technical Decisions, Events, Lessons)
• Scored by importance (0-10) based on keywords and patterns

Phase 3: Consolidation

• Exact dedup — Hash-based removal of identical entries
• Fuzzy dedup — Bigram similarity (Dice coefficient) removes entries >75% similar, keeping the most recent
• Stale pruning — Removes old debugging notes (>14 days), temporary workarounds (>30 days), completed tasks (>7 days)
• Date normalization — Converts "yesterday", "today", "last week" to absolute dates using the file's date as reference
• Protected entries — Entries containing ⚠️, IMPORTANT, NEVER, ALWAYS are never pruned

Phase 4: Prune & Index

Enforces the max line limit by keeping highest-importance + most-recent entries. Writes a clean MEMORY.md with category sections and metadata.

Trigger Conditions

By default, consolidation triggers when both conditions are met:

• 24+ hours since last consolidation
• 5+ new daily files since last consolidation

Use --force to bypass these conditions.

Configuration

Create .autodream.json in your workspace root:

{
  "maxLines": 200,
  "lookbackDays": 30,
  "memoryDir": "memory",
  "memoryIndex": "MEMORY.md",
  "categories": [
    "People & Relationships",
    "Projects & Work",
    "Preferences & Style",
    "Technical Decisions",
    "Important Events",
    "Lessons Learned"
  ],
  "preservePatterns": ["⚠️", "IMPORTANT", "NEVER", "ALWAYS"],
  "triggerThreshold": {
    "minHoursSinceLastRun": 24,
    "minNewFiles": 5
  }
}

Output Example

# Long-Term Memory
<!-- Last consolidated: 2026-03-25T17:00:00Z | Files processed: 29 | Entries: 123 -->

## People & Relationships
- **Bob Smith** — Team Lead test run passed, all scores ≥4/5 (2026-03-25)
- **Carol Davis** — Q2 budget approved (2026-03-21)

## Projects & Work
- **Acme Corp** — Q1 review: 30.1% margin, targeting 40% (2026-03-25)
- **Project Alpha** — Supabase + Vercel architecture decided (2026-03-18)

## Preferences & Style
- Values precision, no tolerance for hallucinations (2026-02-01)
- CET timezone for all scheduling (2026-02-01)

## Technical Decisions
- Redis for session caching (2026-03-21)
- n8n for workflow automation (2026-02-06)

## Important Events
- Q1 2026 engineering review completed (2026-03-25)

## Lessons Learned
- Always test database migrations on staging first (2026-03-22)
- Supabase RLS policies don't apply to service role key (2026-03-22)

Safety

• Non-destructive — Previous MEMORY.md is always backed up to memory/.autodream-backups/
• Reports — Consolidation details saved to memory/.autodream-reports/
• Dry run — Always preview with --dry-run before running
• Protected entries — Configurable patterns that are never pruned
• Zero dependencies — Pure Node.js, no external packages

OpenClaw Skill

The skill/ directory contains an OpenClaw skill definition. Install it in your workspace:

cp -r skill/SKILL.md ~/.openclaw/workspace/skills/autodream/SKILL.md

Or use with heartbeats by adding to your HEARTBEAT.md:

## Memory Consolidation
- Run `autodream <workspace> --stats` to check if needed
- If "Would trigger: ✅ yes", run `autodream <workspace> --verbose`

Comparison: Claude Code Autodream vs OpenClaw Autodream

| Feature | Claude Code | OpenClaw Autodream | |---------|------------|-------------------| | Automatic trigger | ✅ 24h + 5 sessions | ✅ 24h + 5 files | | Session transcript mining | ✅ (LLM-powered) | ❌ (file-based only) | | Date normalization | ✅ | ✅ | | Contradiction resolution | ✅ (LLM-powered) | ✅ (recency-based) | | Deduplication | ✅ | ✅ (exact + fuzzy) | | Stale pruning | ✅ | ✅ | | Size enforcement | ✅ (200 lines) | ✅ (configurable) | | Manual trigger | /dream command | autodream CLI | | Backups | Unknown | ✅ Always | | Reports | Unknown | ✅ JSON reports | | LLM required | Yes | No (pure text) | | Background execution | ✅ | Via cron/heartbeat |

Development

# Run tests
npm test

# Run against a workspace
node bin/autodream.js /path/to/workspace --dry-run --verbose

License

MIT