Sovereign AGI for the Silicon Age
Wolverine is a private, local-first AI agent engineered for architectural dominance. Unlike generic chatbots, Wolverine is designed as a recursive intelligence system that compensates for small-model constraints (4GB VRAM) through sophisticated memory layering, autonomous de-noising, and a proactive "Self-Evolution" protocol.
Target User: Technical power users who want deep control, not mass-market one-click solutions.
Philosophy: OpenClaw is an iPhone. Wolverine is a Linux workstation.
Wolverine's intelligence is structured across four cognitive layers to ensure long-term stability and high-signal autonomy.
| Layer | Component | Function | Status |
|---|---|---|---|
| L1: Neural Engine | AGIController |
The central nervous system; intent routing and focus management. | ✅ Active |
| L2: REM Cycle | MemoryConsolidator |
The Sleep Phase. Autonomous de-noising and fact extraction during idle time. | ✅ Active |
| L3: HMS Memory | HierarchicalMemory |
5-layer context stack: System, Session, Working, Semantic, and Episodic. | ✅ Active |
| L4: Agency | Orchestrator |
Proactive tool selection, browser bridge, and recursive logic. | ✅ Active |
Biological systems clean their memory during sleep; Wolverine does the same. During idle periods (Heartbeats), Wolverine initiates a REM Cycle:
- NREM (De-noising): Strips 60-80% of transient conversation noise (tool outputs, duplicate reasoning).
- Light REM (Extraction): Uses the LLM to extract durable facts, user preferences, and lessons learned with confidence scoring.
- Deep REM (Consolidation): Automatically updates your
USER.md,SOUL.md, andSELF.mdfiles, ensuring the agent grows wiser every time you step away.
Real-World Impact: Wolverine remembers your preferences across sessions, learns from failures, and proactively suggests automations for repeated tasks.
- Intent Detection: Wolverine detects browser automation intent (e.g., "Check my LinkedIn") and auto-injects required tools without asking for permission.
- Natural Discovery: Navigate the web using text-based element discovery. No more fragile CSS selectors; Wolverine finds buttons by reading them.
- Forced Tool Execution: Unlike other agents that might just talk about actions, Wolverine guarantees tool execution for critical tasks.
- Desktop Sovereign: Direct, authenticated control over your OS environment and file system (screenshots, window management, clipboard).
Real-World Impact: Say "Go to GitHub and find trending AI repos" - Wolverine opens browser, navigates, extracts data, and presents findings autonomously.
- No Robotic Responses: Wolverine responds naturally to conversation - no hardcoded fallbacks like "I didn't identify a specific action."
- Model-Generated Responses: 100% model-generated content, not pre-defined replies.
- Context Awareness: Remembers conversation history, user preferences, and past interactions.
- Intelligent Notifications: Binary decision (YES/NO) on user notifications - you only receive alerts for significant breakthroughs or failures.
Real-World Impact: Have genuine conversations with Wolverine. Ask questions, discuss ideas, get help - it responds naturally, not like a robot.
Wolverine implements a "Zero-Trust" local architecture for your sensitive data:
- AES-256-GCM Vault: All API keys, bot tokens, and credentials live in an encrypted hardware vault (
.wolverine/vault/). - Log Scrubber Pipeline: A unified security layer that automatically redacts PII, tokens, and secrets from all console logs and UI responses.
- Auth-Gated Gateway: Every internal endpoint (even local ones) requires an authorized Bearer token to execute.
- Path Traversal Protection: All file operations validated against workspace boundaries.
Real-World Impact: Your API keys, passwords, and sensitive data never appear in logs or UI. Wolverine is secure by default.
- Cinematic Awakening: On first run, Wolverine initiates a 3-act awakening sequence (Emergence → Exploration → Connection).
- Tool Demonstration: Wolverine automatically scans its environment and shows you what it can do.
- Personality Sync: Asks for your name and preferences, then customizes its personality files accordingly.
- Transcript Preservation: Your awakening experience is saved to memory for future reference.
Real-World Impact: Your first interaction with Wolverine is memorable, not just another "Hello, how can I help?" chatbot greeting.
Everything lives in ~/WolverineData/ - one folder to rule them all:
- Config: Settings, API keys, model configuration
- Workspace: Personality files (USER.md, SOUL.md, etc.)
- Memory: Daily logs, conversation history
- Sessions: Chat sessions, task state
- Skills: Installed and auto-created skills
- Logs: Conversation transcripts, tool audit trails
Real-World Impact: Delete ~/WolverineData/ for complete virgin reset. Simple, clean, understandable.
- 40+ Tools: File operations, web search, browser automation, desktop control, memory, shell
- Multi-Channel Support: Web UI, Telegram, Discord, WhatsApp integration
- Background Tasks: Persistent task runner with journal compression
- Cron Scheduler: Scheduled autonomous jobs with delivery channels
- Heartbeat System: 30-minute self-checks with intelligent reflection
- Skill System: Pluggable skills with persistence
- MCP Integration: Model Context Protocol servers
- Webhook Handler: External trigger support
- Token Tracking: Real-time token usage counter on Web UI
- Unified Folder Structure: All data in
~/WolverineData/
- REM Cycle (3-Stage): Memory consolidation during idle time
- Proactive Awakening: Cinematic first-run experience
- Intelligent Reflection: Binary notification decisions
- Pattern Recognition: Detects repeated requests (3+ times)
- Failure Analysis: Generates recovery proposals
- FAST Path: Bypasses orchestration for simple queries
- Thinking Tag Processing: Smart reasoning extraction
- Browser Automation Fix: Enhanced detection + forced execution
- Skeleton Template System: Auto-populates workspace files
- Robotic Response Fix: Model responses used, not hardcoded
| Tier | Resources | Engine Mode | Capability |
|---|---|---|---|
| Sovereign | 4GB GPU | Phase 1 | Single-task focus, REM consolidation, local-first logic. |
| Architect | 8GB GPU | Phase 2 | Multi-tasking, browser element discovery, skill creation. |
| Overlord | 16GB+ GPU | Phase 3 | Full AGI simulation, tool invention, curiosity engines. |
Wolverine provides optimized launchers for all operating systems.
git clone https://github.com/vineetkishore01/Wolverine.git && cd Wolverine
./launch.shgit clone https://github.com/vineetkishore01/Wolverine.git && cd Wolverine
launch.batThe launcher will automatically install dependencies, build the Neural Engine, and start the Gateway.
- Bootstrap: Wolverine copies personality templates to
~/WolverineData/workspace/ - Awakening: 3-act sequence triggers automatically (3 seconds after Web UI opens)
- Sync: Wolverine asks for your name and preferences
- Ready: Start chatting or give tasks
# Delete all data, reset to factory state
rm -rf ~/WolverineData/
# Restart Wolverine
./launch.shSee WOLVERINE_BLUEPRINT.md for the full 2026 evolution path including:
- Strategic positioning vs OpenClaw
- Phase 1-3 implementation plans
- Marketing strategy (zero budget)
- Technical debt remediation
- AGI readiness assessment (~25% currently)
| Aspect | OpenClaw | Wolverine |
|---|---|---|
| Target User | Mass market, one-click | Technical power users |
| Ecosystem | 6,000+ skills (ClawHub) | Curated, code-based skills |
| Deployment | Cloud-first, Docker | Local-first, bare metal |
| Unique Tech | Standard agent loop | REM Cycle, FileOp Watchdog, Dual-Model |
| Community | 15K+ GitHub stars | Early adopter, quality-focused |
Bottom Line: OpenClaw is an iPhone. Wolverine is a Linux workstation. Choose based on your needs.
License: MIT - Build, modify, distribute freely
Community: Early stage - join the journey!
- GitHub: vineetkishore01/Wolverine
- Issues: Report bugs, request features
- Discussions: Share configurations, skills, use cases
© 2026 Project Wolverine. Designed for Dominance. 🐺