aradotso/skills/openclaw-auto-dream

---
name: openclaw-auto-dream
description: Automatic cognitive memory consolidation for OpenClaw/MyClaw agents — sleep cycles, importance scoring, forgetting curves, knowledge graphs, and health dashboards.
triggers:
  - set up auto-dream memory for my openclaw agent
  - configure memory consolidation for myclaw
  - how do I install openclaw auto-dream
  - my ai agent keeps forgetting things between sessions
  - set up dream cycles for memory management
  - how does auto-dream importance scoring work
  - export or import memory bundle between myclaw instances
  - show me the memory health dashboard
---

# OpenClaw Auto-Dream

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.

OpenClaw Auto-Dream is a cognitive memory architecture for [OpenClaw](https://github.com/openclaw/openclaw) agents (as used on [MyClaw.ai](https://myclaw.ai)). It gives your AI agent the ability to periodically "sleep and dream" — scanning raw daily logs, extracting structured knowledge, scoring importance, applying forgetting curves, building a knowledge graph, and surfacing non-obvious insights. The result is an agent that genuinely learns and connects the dots over time rather than accumulating stale, disconnected files.

---

## How It Works

Auto-Dream runs a **three-phase dream cycle** (default: 4 AM daily via cron):

1. **Collect** — Scans unconsolidated daily logs (last 7 days), detects priority markers, extracts decisions/people/facts/projects/lessons/procedures/open threads.
2. **Consolidate** — Routes each insight to one of five memory layers, deduplicates semantically, assigns unique IDs (`mem_NNN`), creates relation links.
3. **Evaluate** — Scores importance, applies forgetting curves, calculates a 5-metric health score, generates insights, writes dream report, sends notification.

### Five Memory Layers

| Layer | Storage | Purpose |
|-------|---------|---------|
| Working | LCM plugin (auto-detected) | Real-time context compression & semantic recall |
| Episodic | `memory/episodes/*.md` | Project narratives, event timelines |
| Long-term | `MEMORY.md` | Facts, decisions, people, milestones, strategy |
| Procedural | `memory/procedures.md` | Workflows, preferences, tool patterns |
| Index | `memory/index.json` | Metadata, scores, relations, health stats |

---

## Installation

### Via ClawHub (Recommended)

Tell your MyClaw agent:

Install the openclaw-auto-dream skill from ClawHub

Or manually inside your OpenClaw agent environment:

```bash
claw skill install openclaw-auto-dream

Manual Installation

Clone into your OpenClaw skills directory:

git clone https://github.com/LeoYeAI/openclaw-auto-dream.git \
  ~/.openclaw/skills/openclaw-auto-dream

Then register the skill with your agent:

claw skill register ~/.openclaw/skills/openclaw-auto-dream

First-Time Setup

After installation, tell your agent:

Set up auto-dream

The setup wizard will:

• Detect whether the optional LCM plugin is available (for the Working memory layer)
• Create the memory/ directory structure
• Initialize memory/index.json with default health metrics
• Ask for your preferred notification level (silent / summary / full)
• Schedule the dream cron job (default: 0 4 * * *)

---

Configuration

Auto-Dream is configured via memory/config.json in your agent's workspace:

{
  "dream_schedule": "0 4 * * *",
  "notification_level": "summary",
  "scan_window_days": 7,
  "forgetting": {
    "min_age_days": 90,
    "importance_threshold": 0.3
  },
  "scoring": {
    "recency_half_life_days": 180,
    "permanent_marker": "⚠️ PERMANENT",
    "high_marker": "🔥 HIGH",
    "pin_marker": "📌 PIN"
  },
  "layers": {
    "working_lcm": true,
    "episodic": true,
    "long_term": true,
    "procedural": true
  },
  "export": {
    "compress": true,
    "include_archive": false
  }
}

Environment Variables

If you need to override config values via the environment (e.g. in CI or multi-instance deployments):

AUTODREAM_SCHEDULE="0 2 * * *"          # Override cron schedule
AUTODREAM_NOTIFY_LEVEL="full"           # silent | summary | full
AUTODREAM_SCAN_DAYS=14                  # Days of logs to scan per cycle
AUTODREAM_FORGET_THRESHOLD=0.25         # Importance below which entries are archived

---

Key Commands (Agent Natural Language)

These phrases trigger built-in Auto-Dream intents inside your OpenClaw agent:

| Phrase | Action | |--------|--------| | "Dream now" | Trigger an immediate full dream cycle | | "Show memory dashboard" | Generate and open the interactive HTML dashboard | | "What do you remember about [topic]?" | Semantic search across all memory layers | | "Memory health" | Print current 5-metric health score | | "Export memory bundle" | Export all layers to memory/export-YYYY-MM-DD.json | | "Import memory bundle" | Merge an exported bundle into current memory | | "Export only procedures" | Selective single-layer export | | "Forget [topic]" | Immediately archive entries matching topic | | "Pin this" | Mark current context with 📌 PIN (immune to forgetting) | | "What did you learn last week?" | Show insights from the last 7 dream logs |

---

Priority Markers in Daily Logs

Auto-Dream scans your agent's daily log files for special markers during the Collect phase. Use these in any log entry or conversation note:

<!-- important -->
Decided to use Postgres over SQLite for the user DB — scalability concern.

⚠️ PERMANENT
Client prefers all reports in US Letter format, not A4.

🔥 HIGH
The deploy pipeline breaks when NODE_ENV is not explicitly set.

📌 PIN
Weekly sync with Alex every Tuesday at 10 AM.

| Marker | Effect | |--------|--------| | <!-- important --> | Extracted and routed to appropriate memory layer | | ⚠️ PERMANENT | Always scores 1.0 importance; never archived | | 🔥 HIGH | Base weight doubled during importance scoring | | 📌 PIN | Immune to forgetting curve; always retained |

---

Importance Scoring

Every memory entry is scored on each dream cycle:

importance = (base_weight × recency_factor × reference_boost) / 8.0

Where:

• recency_factor = max(0.1, 1.0 - days_since_created / 180)
• reference_boost = log₂(reference_count + 1)
• base_weight doubles for 🔥 HIGH entries; ⚠️ PERMANENT always returns 1.0

Example (Python-style pseudocode)

import math

def score_entry(entry: dict, today_date) -> float:
    if "⚠️ PERMANENT" in entry.get("markers", []):
        return 1.0

    days_old = (today_date - entry["created_at"]).days
    recency = max(0.1, 1.0 - days_old / 180)

    refs = entry.get("reference_count", 0)
    ref_boost = math.log2(refs + 1) if refs > 0 else 1.0

    base = entry.get("base_weight", 1.0)
    if "🔥 HIGH" in entry.get("markers", []):
        base *= 2.0

    return min(1.0, (base * recency * ref_boost) / 8.0)

---

Forgetting Curve & Archival

Entries are never deleted — only gracefully archived:

def should_archive(entry: dict, today_date) -> bool:
    # Immune markers
    immune = {"⚠️ PERMANENT", "📌 PIN"}
    if immune & set(entry.get("markers", [])):
        return False

    days_unreferenced = (today_date - entry["last_referenced"]).days
    importance = entry["importance_score"]

    return days_unreferenced > 90 and importance < 0.3

def archive_entry(entry: dict):
    summary = f"[{entry['id']}] {entry['title']} — {entry['one_line_summary']}"
    append_to_file("memory/archive.md", summary)
    entry["status"] = "archived"
    # Original ID preserved for relation tracking

---

Health Score

health = (
    freshness   × 0.25 +
    coverage    × 0.25 +
    coherence   × 0.20 +
    efficiency  × 0.15 +
    reachability× 0.15
) × 100

| Metric | Definition | |--------|-----------| | Freshness | % of entries referenced in the last 30 days | | Coverage | % of knowledge categories updated in the last 14 days | | Coherence | % of entries with at least one relation link | | Efficiency | Inversely proportional to MEMORY.md line count (penalises bloat) | | Reachability | Knowledge graph connectivity via union-find across entry relations |

Check health anytime:

"Memory health"
→ 🩺 Health: 79/100
   Freshness: 0.81  Coverage: 0.74  Coherence: 0.68
   Efficiency: 0.91  Reachability: 0.72
   ⚠️ Coherence declining — consider linking isolated entries

---

Cross-Instance Memory Migration

Export from one MyClaw instance:

"Export memory bundle"
→ Saved: memory/export-2026-03-28.json (142 entries, 3 layers)

The export format:

{
  "version": "4.0",
  "exported_at": "2026-03-28T06:00:00Z",
  "source_instance": "myclaw-instance-abc123",
  "layers": {
    "long_term": [ /* entries */ ],
    "procedural": [ /* entries */ ],
    "episodic": [ /* entries */ ]
  },
  "index": { /* metadata + scores */ }
}

Import on another instance (newer entry wins on conflict):

"Import memory bundle"
→ Upload or paste path to export JSON
→ Merging 142 entries... 138 new, 4 conflicts resolved (newer wins)
→ Pre-import backup saved to memory/pre-import-backup-2026-03-28.json

Selective export:

"Export only procedures"
→ Saved: memory/export-procedures-2026-03-28.json (23 entries)

---

Memory File Structure

After setup, your agent workspace will contain:

workspace/
├── MEMORY.md                        # Long-term memory layer (human-readable)
├── memory/
│   ├── config.json                  # Auto-Dream configuration
│   ├── index.json                   # Entry metadata, scores, relations, health
│   ├── procedures.md                # Procedural memory layer
│   ├── archive.md                   # Archived (forgotten) entries — one-liners
│   ├── dream-log.md                 # Appended report after each cycle
│   ├── dashboard.html               # Generated interactive dashboard
│   ├── episodes/
│   │   ├── project-alpha.md
│   │   └── onboarding-2026-01.md
│   └── exports/
│       └── export-2026-03-28.json
└── logs/
    ├── daily-2026-03-28.md          # Raw daily logs (scanned by Auto-Dream)
    └── daily-2026-03-27.md

---

Interactive Dashboard

Generate the zero-dependency HTML dashboard:

"Show memory dashboard"
→ Generated: memory/dashboard.html

The dashboard includes:

• Animated health gauge (0–100)
• 5 metric cards with colour-coded status (green ≥ 0.8, amber ≥ 0.6, red < 0.6)
• Memory distribution donut chart (entries per layer)
• Importance histogram (score distribution across all entries)
• Health trend line chart (last 30 dream cycles)
• Force-directed knowledge graph (linked entries visualised)
• Recent changes, dream insights, stale entry warnings

Open it in any browser — no server required, all data embedded inline.

---

Dream Log Format

After each cycle, a report is appended to memory/dream-log.md:

## Dream Cycle — 2026-03-28T04:00:12Z

**Health:** 82/100 (↑3 from last cycle)
**Entries:** +5 new · ~3 updated · -1 archived
**Duration:** 14.2s

### New Memories
- mem_087: Decided to migrate auth to Clerk (episodic + long-term)
- mem_088: Alex prefers async updates over meetings (long-term)

### Archived
- mem_034: [compressed] Early Stripe test key config — 2025-09-01

### Insights
1. Project Beta's scope decisions mirror what caused delays in Project Alpha — pattern detected.
2. Strategic decisions cluster on Mondays — likely tied to weekly planning sessions.
3. No lessons recorded for last 3 completed projects — retrospective habit may be slipping.

### Suggestions
- Link mem_072 (auth strategy) ↔ mem_087 (Clerk migration) — high semantic overlap
- mem_091 unreferenced for 78 days — consider pinning or letting it expire

---

Notification Levels

Configure in memory/config.json or tell your agent "Set notifications to [level]":

| Level | What You Receive | |-------|-----------------| | silent | Nothing pushed — logged to dream-log.md only | | summary | 🌀 Health: 82/100 \| +5 new, ~3 updated, -1 archived \| 💡 Top insight | | full | Complete dream report with all sections pushed to your chat channel |

Push is delivered to whichever channel your MyClaw agent is connected to (Telegram, Discord, Slack, WhatsApp, etc.).

---

Troubleshooting

Dream cycle not running

# Check cron is registered
crontab -l | grep autodream

# Run a manual cycle and watch output
claw skill run openclaw-auto-dream --trigger dream --verbose

# Check for log errors
tail -50 ~/.openclaw/logs/skills.log | grep auto-dream

memory/index.json is corrupted or missing

"Rebuild memory index"
→ Scans all memory layer files and regenerates index.json from scratch
→ Note: reference counts and relation links will be reset

Health score stuck low despite active use

Common causes and fixes:

| Symptom | Cause | Fix | |---------|-------|-----| | Coherence < 0.5 | Many unlinked entries | "Link related memories" — agent will suggest and apply links | | Freshness < 0.5 | Old entries not being referenced | "What do you remember about X?" — queries boost reference counts | | Efficiency < 0.5 | MEMORY.md has grown too large | "Compress long-term memory" — consolidates verbose entries | | Reachability < 0.5 | Isolated knowledge clusters | Check dream insights — they will flag disconnected subgraphs |

LCM plugin not detected (Working layer disabled)

"Check working memory status"
→ ⚠️ LCM plugin not found — Working layer disabled
→ Install with: claw plugin install lcm
→ Then: claw skill reconfigure openclaw-auto-dream

Entries marked PERMANENT are being archived

This is a bug — ⚠️ PERMANENT entries should never be archived. Force a fix:

# In your agent workspace
claw skill run openclaw-auto-dream --repair-permanent-markers

---

Common Patterns

Mark important decisions automatically in daily logs

Add to your agent's daily log template in memory/procedures.md:

## Decision Template
When recording a decision, always include:
🔥 HIGH
Decision: [what was decided]
Rationale: [why]
Alternatives considered: [what else was evaluated]

Trigger a dream before a big project review

"Dream now, then show me everything you remember about Project Alpha"

Weekly memory hygiene check

"Memory health"
"What's stale?"
"Link related memories"
"Dream now"

Multi-instance workflow (e.g. work vs. personal MyClaw)

# On work instance
"Export only procedures"
# Copy memory/export-procedures-2026-03-28.json to personal instance
# On personal instance
"Import memory bundle"
# Select the procedures export — workflows sync across instances