memory-hybrid

<div align="center">

🦁 LeOpus

Battle-tested AI assistant. One command. One token. Zero cost.

A pre-configured OpenClaw setup with a 5-agent hierarchy, task lifecycle management, and execution discipline — running entirely on local hardware.


</div>

---

What You Get

• 5-agent hierarchy — Main orchestrator + Builder, Scout, Operator, Analyst
• Local LLM — Qwen2.5 7B via Ollama. No API keys. No monthly bills. Ever.
• Telegram interface — Talk to your AI from your phone
• Task lifecycle — Every multi-step job gets tracked, checkpointed, and verified
• Compaction survival — Agent recovers context after memory resets
• Heartbeat monitoring — Periodic system health checks
• Execution discipline — No hallucinated completions. Proof or it didn't happen.

Requirements

| Component | Minimum | Recommended | |-----------|---------|-------------| | RAM | 16 GB | 32 GB+ | | GPU | None (CPU works) | NVIDIA or Apple Silicon | | OS | macOS or Ubuntu/Debian | Any with bash + Node.js | | Disk | 10 GB free | 20 GB free |

Quick Start

git clone https://github.com/rodawg71/leopus.git
cd leopus
./setup-leopus.sh

The installer handles everything:

1. Hardware check — validates RAM and detects GPU 2. Node.js — installs v22+ if missing 3. Ollama — installs and pulls Qwen2.5 7B (~5GB download) 4. OpenClaw — installs via npm 5. Your token — paste your Telegram bot token (the only input) 6. Deploy — copies configs, generates secrets, starts the gateway

Then open Telegram and message your bot.

Get a Telegram Bot Token

1. Open Telegram → search @BotFather 2. Send /newbot 3. Choose a name and username 4. Copy the token

That's it. That's the only thing you need.

Architecture

You (Telegram) → LeOpus (Main Agent)
                    ├── Builder   — writes code, creates files
                    ├── Scout     — researches, fetches info
                    ├── Operator  — browser automation, system ops
                    └── Analyst   — data analysis, summaries

LeOpus (main) never executes directly — it plans, delegates to sub-agents, and verifies results. Think CEO, not intern.

Key Files

After install, your workspace lives at ~/.openclaw/workspace/:

| File | Purpose | |------|---------| | SOUL.md | Agent personality, values, execution rules | | AGENTS.md | Boot sequence, task lifecycle, write discipline | | USER.md | Info about you (name, timezone) | | IDENTITY.md | Agent identity | | MEMORY.md | Memory index | | HEARTBEAT.md | Automated health check schedule | | templates/task-template.md | Template for multi-step tasks |

Everything is plain Markdown. Edit any file to customize behavior.

Commands

openclaw gateway status     # check if running
openclaw gateway restart    # restart
openclaw gateway stop       # stop
openclaw gateway logs       # view logs

What Makes This Different

Most AI setups are a model behind an API. LeOpus is an opinionated system built from months of real-world usage:

• Truth over agreement — it will push back if you're wrong
• Proof rule — every claim of completion includes evidence
• Write discipline — all file writes are chunked and verified
• Loop detection — catches itself if stuck in reasoning loops
• No completion theater — never says "done" without proof

Customization

Change the LLM

Edit ~/.openclaw/openclaw.json — swap qwen2.5:7b for any Ollama model:

ollama pull llama3:8b          # Meta Llama 3
ollama pull mistral:7b         # Mistral
ollama pull deepseek-v3:8b     # DeepSeek

Or point to a cloud provider (OpenAI, Anthropic, etc.) by updating the

providers section in openclaw.json.

Add Skills

openclaw skills install <skill-name>

Browse available skills at clawhub.com.

Change Personality

Edit ~/.openclaw/workspace/SOUL.md. It's just Markdown. Want a friendlier assistant? A stricter one? A pirate? Change the file.

FAQ

Q: Is this really free? Yes. Ollama is open source. Qwen2.5 is open weight. OpenClaw is open source. The only cost is the electricity to run your machine.

Q: How much VRAM does it need? Qwen2.5 7B needs ~5GB. Works on most NVIDIA cards (GTX 1060+) and all Apple Silicon. CPU-only works too, just slower.

Q: Can I use a cloud LLM instead? Yes. Edit openclaw.json to add OpenAI, Anthropic, or any OpenAI-compatible API.

Q: What about privacy? Everything runs locally by default. Your conversations never leave your machine unless you configure a cloud provider.

Credits

Built by @rodawg71, evolved through daily use with OpenClaw.

Based on clawdboss by NanoFlow-io.

License

MIT