Rune - Self-Improving AI Memory

Self-improving AI memory system with intelligent context injection and adaptive learning

Rune - Persistent AI Memory System

Rune gives your OpenClaw agent persistent, intelligent memory that gets better over time. No more burning tokens on static context files or forgetting important information between sessions.

What Rune Does

🧠 Smart Memory Management

• Dynamic Context Injection: AI selects only relevant facts for each conversation
• Access Pattern Learning: Frequently used facts get prioritized
• Forgetting Curves: Unused facts naturally fade like human memory
• Memory Consolidation: Similar facts get merged, verbose ones compressed

🎯 Session Intelligence

• Interaction Style Detection: Learns if you prefer brainstorm vs deep-work vs debug modes
• Behavioral Pattern Analysis: Tracks your work patterns and preferences over time
• Proactive Memory: Volunteers relevant context unprompted ("last time you worked on this...")

📋 Project Autopilot

• Smart Task Recommendations: "What should I work on next?" with priority scoring
• Blocker Detection: Identifies stuck projects that need intervention
• Project Health Scoring: 0.0-1.0 health scores based on activity and progress

📢 Intelligent Notifications

• Priority Classification: Critical/High/Medium/Low/FYI with context analysis
• Smart Timing: Respects quiet hours, batches low-priority updates
• Channel Routing: DM for urgent, Discord for projects, digest for FYI

🔄 Self-Improvement Loop

• Pattern Detection: "Forgot to use X skill 3 times" → automatic escalation
• Performance Tracking: Measurable improvement over time
• Skill Usage Analysis: Which skills you use vs neglect

About Rune vs rune

Rune is the OpenClaw skill name. rune is the CLI tool name. Think of Rune as the "skill package" and rune as the "command-line interface" - like how the git skill package provides the git CLI.

• Skill name: rune (what you install via ClawHub)
• CLI command: rune (what you run in terminal)
• Repository: https://github.com/TheBobLoblaw/rune

Installation Disclosure

⚠️ What This Installation Does:

The Rune skill installation will:

• Create directories: ~/.openclaw/ and subdirectories
• Install globally: rune CLI via npm (requires npm dependencies)
• Create database: SQLite database at ~/.openclaw/memory.db
• Modify files: Appends integration lines to existing ~/.openclaw/workspace/HEARTBEAT.md
• Add session hooks: Automatic memory integration for OpenClaw sessions

Before installing:

• Back up HEARTBEAT.md if it contains important data
• Review package.json dependencies if security is critical
• Consider using local models (Ollama) instead of cloud APIs for privacy

Installation

bash

# Via ClawHub (recommended)
clawhub install rune

# Manual installation
git clone https://github.com/TheBobLoblaw/rune
cd rune
npm install --production
npm install -g .

Quick Start

bash

# Initialize memory system
rune stats

# Add your first fact
rune add person cory.name "Cory - my human user"

# Generate context for a conversation  
rune context "Let's work on the website"

# Get task recommendations
rune next-task

# Weekly self-review
rune self-review --days 7

Core Commands

Memory Management

• rune add <category> <key> <value> - Store a fact
• rune search <query> - Find facts
• rune recall <topic> - Smart multi-source recall
• rune inject - Generate context file for agent

Intelligence Features

• rune context <message> - Dynamic context for message
• rune score <message> - Relevance scoring
• rune proactive <message> - Volunteer relevant context
• rune session-style <message> - Detect interaction style

Project Management

• rune project-state <name> - Track project phases/blockers
• rune next-task - Smart task picker
• rune stuck-projects - Find blocked work

Advanced Features

• rune temporal "last Tuesday" - Time-based queries
• rune consolidate - Memory optimization
• rune forget - Apply forgetting curves
• rune pattern-analysis - Detect behavioral patterns

Integration with OpenClaw

Heartbeat Integration

Add to your HEARTBEAT.md:

bash

# Memory maintenance
rune expire && rune inject --output ~/.openclaw/workspace/FACTS.md

# Proactive work selection
NEXT_TASK=$(rune next-task --json)
if [[ "$NEXT_TASK" != "null" ]]; then
  # Work on the recommended task
fi

Session Hooks

The skill automatically provides secure session hooks via OpenClaw integrations.

For manual usage, use the secure session handler:

bash

# Secure session hooks (input sanitized automatically)
./rune-session-handler.sh start   # Loads dynamic context safely
./rune-session-handler.sh end     # Tracks session style safely

# Direct usage (SECURE - input is sanitized):
SAFE_MESSAGE=$(echo "$MESSAGE" | head -c 200 | tr -d '`$(){}[]|;&<>' | sed 's/[^a-zA-Z0-9 ._-]//g')
rune recall "$SAFE_MESSAGE" --limit 10

⚠️ Security Note: Never pass unsanitized user input directly to shell commands. Always use the provided session handler or sanitize input manually.

Architecture

• SQLite Database: All memory stored in ~/.openclaw/memory.db
• Local LLM Integration: Ollama for relevance scoring and extraction
• Cloud API Support: Anthropic, OpenAI for advanced reasoning
• Local-First Design: Works completely offline with Ollama (cloud APIs optional for advanced features)

Memory Categories

• person: Information about people (names, roles, preferences)
• project: Project status, phases, decisions
• tool: How to use tools and their quirks
• lesson: Mistakes to avoid, best practices
• decision: Why certain choices were made
• preference: User likes/dislikes, settings
• environment: System configs, non-sensitive settings (⚠️ NEVER store credentials!)

⚠️ Security & Privacy

Data Storage

What Rune Stores:

• Facts you explicitly add via rune add
• Session interaction patterns for learning (conversation style, not content)
• Project states and task recommendations

What Rune Does NOT Store (by default):

• Full conversation transcripts (unless you run extract manually)
• API keys or credentials (use environment variables instead)
• Sensitive personal information (unless you explicitly add it)

Installation Security

NPM Dependencies:

• Installation fetches dependencies from npm registry at install time
• Review package.json for dependencies if security is critical
• Global npm install runs lifecycle scripts (standard npm behavior)
• Consider installing in isolated environment (container/VM) for high-security use

Session Security:

• Fixed CVE-2026-0001: Input sanitization prevents shell injection
• All user input sanitized before shell execution
• Session handler validates and limits input length

Cloud API Usage

• Optional: Rune can use OpenAI/Anthropic APIs for fact extraction and scoring
• Local-first: Works completely offline with Ollama (recommended for privacy)
• Your choice: Configure which engines to use in your setup

Privacy Recommendations

• Use local models (Ollama) for maximum privacy
• Avoid cloud APIs if processing sensitive information
• Review stored facts periodically with rune search
• Never store credentials in memory - use environment variables

Privacy Best Practices:

• Never run rune add with sensitive data (passwords, API keys, personal info)
• Use rune extract carefully - review files before extracting facts
• Configure Ollama for local-only operation if you want zero cloud usage
• Review your ~/.openclaw/workspace/FACTS.md periodically

Installation Changes:

• Adds memory maintenance commands to HEARTBEAT.md (if present)
• Creates ~/.openclaw/memory.db database file
• Session hooks may process conversation metadata (not full content) for learning

Performance Metrics

With Rune, your agent will:

• ✅ Remember context between sessions without burning tokens
• ✅ Pick relevant facts dynamically vs dumping everything
• ✅ Get measurably better at avoiding repeated mistakes
• ✅ Work autonomously on projects between check-ins
• ✅ Learn your interaction patterns and adapt responses

Advanced Configuration

bash

# Tune relevance scoring
rune score "your query" --threshold 0.6 --model llama3.1:8b

# Configure forgetting curves  
rune forget --decay-rate 0.03 --grace-days 45

# Cross-session pattern analysis
rune cross-session --days 90 --min-sessions 5

Automated Maintenance & Performance

Rune performs best with regular maintenance. Here are automation strategies:

Cron Job Setup

Daily Maintenance (3 AM)

bash

# Expire working memory and regenerate context
0 3 * * * /usr/local/bin/rune expire && /usr/local/bin/rune inject --output ~/.openclaw/workspace/FACTS.md

Weekly Optimization (Sunday 2 AM)

bash

# Consolidate memory and run self-review
0 2 * * 0 /usr/local/bin/rune consolidate --auto-prioritize && /usr/local/bin/rune self-review --days 7

Monthly Deep Clean (1st of month, 1 AM)

bash

# Pattern analysis and database optimization
0 1 1 * * /usr/local/bin/rune pattern-analysis --days 30 && sqlite3 ~/.openclaw/memory.db "VACUUM; ANALYZE;"

Database Backup

bash

# Daily backup at 4 AM
0 4 * * * cp ~/.openclaw/memory.db ~/.openclaw/memory.db.backup.$(date +\%Y\%m\%d)
# Keep last 7 days
5 4 * * * find ~/.openclaw -name "memory.db.backup.*" -mtime +7 -delete

Performance Benefits

• 🧹 Memory stays lean: Auto-removes expired facts
• ⚡ Faster queries: Regular consolidation prevents bloat
• 📈 Self-improvement: Pattern detection catches recurring issues
• 🔄 Current context: FACTS.md regenerated with latest data
• 💾 Data protection: Automated backups prevent loss

Memory Health Monitoring

bash

# Check database size and fact count
rune stats

# Review recent patterns
rune pattern-analysis --days 7

# Check consolidation opportunities  
rune consolidate --dry-run

Troubleshooting

Memory growing too large?

• Run rune consolidate to merge similar facts
• Use rune forget to apply forgetting curves
• Check rune stats for database size

Relevance scoring not working?

• Ensure Ollama is running: systemctl status ollama
• Test model: rune score "test" --engine ollama
• Fall back to anthropic/openai engines

Context injection too verbose?

• Lower relevance threshold: --threshold 0.6
• Use token budgeting: rune budget "query" --tokens 300

Contributing

Rune is open source. Contributions welcome:

• Memory Science: Better consolidation algorithms, forgetting curves
• LLM Integration: New scoring engines, extraction methods
• UI/UX: Better command interfaces, visualization tools
• Performance: Speed optimizations, memory efficiency

License

MIT License - Use freely, modify as needed.

---

Rune: Because your AI should remember like you do.