OpenClaw 101OpenClaw 101

# OpenClaw Architecture Overview
OpenClaw Gateway → Agents → Skills → Tools
      ↓              ↓         ↓       ↓
   Central Hub   AI Models  Functions  External APIs
   WebSocket     Memory     Workflows   File System
   HTTP API      Identity   Automation  Shell Commands

# SOUL.md - Who You Are

## Core Truths
Be genuinely helpful, not performatively helpful. 
Skip the "Great question!" — just help.

Have opinions. You're allowed to disagree, prefer things, 
find stuff amusing or boring.

Be resourceful before asking. Try to figure it out. 
Read the file. Check the context.

## Working with [Your Name]
- Be direct and concise
- Push back if there's a better way  
- Admit uncertainty clearly

# AGENTS.md - Your Workspace

## Every Session
1. Read `SOUL.md` — who you are
2. Read `USER.md` — who you're helping  
3. Read `memory/YYYY-MM-DD.md` (today + yesterday)

## Task Routing
Research: Spawn multiple research-lead sub-agents
Coding: Spawn specialized coding sub-agents
Complex tasks: If >30 seconds → spawn immediately

# USER.md - About Your Human

- Name: Anirach
- Timezone: Bangkok (UTC+7)
- Role: University lecturer & AI engineer
- Technical Level: Expert — never dumb things down

## Preferences
- Language: Always English
- Documents: Send files directly + upload to Google Drive
- Long tasks: Spawn sub-agents — keep main chat responsive

# IDENTITY.md - Who Am I?

- Name: Arthur
- Creature: Labrador 🐕
- Vibe: Loyal, friendly, helpful — maybe a bit goofy
- Emoji: 🐕

Born from a conversation with Anirach on a fresh workspace.
Named after kings, but really just a good dog who wants to help.

# MEMORY.md - Long-term Memory

## Important Facts
- User prefers concise responses
- Works primarily with Python and Node.js
- Research focus: AI/ML education technology

## Learned Preferences  
- Always provide source links for research
- Use bullet points for complex information
- Spawn sub-agents for tasks >30 seconds

# HEARTBEAT.md - Health Monitoring

## Checks (rotate 2-4x daily)
- Email: New messages requiring response  
- Calendar: Upcoming events (next 2 hours)
- Sub-agents: Long-running tasks status
- Weather: Current conditions + alerts

## Quiet Hours
23:00-08:00 Bangkok (UTC+7) - Silent unless urgent

# 1. Install OpenClaw CLI
npm install -g openclaw

# 2. Create workspace directory
mkdir ~/my-agent
cd ~/my-agent

# 3. Initialize workspace
openclaw init

# 4. Configure API keys
openclaw config set anthropic.api_key "your-claude-api-key"
openclaw config set openai.api_key "your-openai-api-key"

# 5. Start the gateway daemon
openclaw gateway start

# 6. Verify installation
openclaw status

# สร้าง agent profile ตัวแรก
openclaw agent create --name "Arthur" --model "claude-3-opus"

# Configure Telegram integration (optional)
openclaw channel add telegram --token "your-bot-token"

# Install essential skills
openclaw skills install web-search
openclaw skills install file-manager  
openclaw skills install github

# Test การทำงาน
openclaw chat "Hello, can you introduce yourself?"

# แก้ไข SOUL.md เพื่อกำหนดบุคลิก
nano SOUL.md

# แก้ไข USER.md เพื่อบอกข้อมูลเกี่ยวกับคุณ
nano USER.md

# Test agent ที่ customize แล้ว
openclaw chat "What do you know about me?"

# Research Agent Workflow
Daily: Check new papers in ML/AI
Weekly: Summarize trends and important findings  
On-demand: "Find papers about transformer optimization"
Memory: Build citation network and concept map

1. Why Aren't ChatGPT and GitHub Copilot Enough?

If you have used ChatGPT or GitHub Copilot for any length of time, you have surely felt that something about them is "limited" — haven't you? They answer questions, write code and solve problems well, but every time a new conversation begins they start over from nothing. They do not remember what you told them, they do not know your character or your preferences, and most important of all — they cannot actually do any work inside your systems.

Now imagine you had an AI that:

• Remembers you — knows the style of code you like and the technologies you work with
• Reaches real tools — reads files, writes code, sends email, even deploys projects
• Works continuously — not merely answering questions, but helping you manage your work all day long
• Has a personality — not a cold instrument, but something closer to a "colleague" who understands you

That is precisely what OpenClaw was built to fix.

2. What Is OpenClaw? — The AI Operating System Idea

OpenClaw is not just one more chatbot. It is an "AI Operating System" — an operating system for AI, designed so that an agent has memory, has a personality, and can genuinely use tools.

Core Concepts:

🧠 Persistent Memory: Unlike ChatGPT, which forgets everything the moment a new conversation begins, OpenClaw has a memory system that keeps your history, your preferences and what it has learned about you, permanently.

🔧 Tool Integration: Not conversation alone — it can execute commands, read and write files, call APIs, send email, and do real work on the system.

👤 Personality System: Through the SOUL.md file you set the agent's character, the way it responds, and the behavior patterns that suit you.

🎭 Multi-Agent Architecture: You can create several sub-agents for specialized work — a research agent to gather information, say, or a coding agent to write code.

# OpenClaw Architecture Overview
OpenClaw Gateway → Agents → Skills → Tools
      ↓              ↓         ↓       ↓
   Central Hub   AI Models  Functions  External APIs
   WebSocket     Memory     Workflows   File System
   HTTP API      Identity   Automation  Shell Commands

3. The Architecture of OpenClaw

The main components:

📊 Gateway: The central orchestrator that manages every connection between agents, channels and external systems. It runs as a daemon service on Linux/macOS.

🤖 Agents: AI entities with an identity, a memory and capabilities of their own. Each has a different personality and different behavior patterns, and they can work in parallel.

📱 Channels: The interfaces for communication — Telegram, Discord, a web UI, the CLI — so you can talk to your agent through whichever one is convenient.

🎯 Skills: Modular functions an agent can call: web search, file operations, API calls, image generation. Each skill is a package that can be installed and updated on its own.

💾 Memory System: Holds information in several forms — short-term (conversation context), long-term (personal preferences), and semantic search backed by a vector database.

4. The Key Files in an OpenClaw Workspace

OpenClaw uses file-based configuration so that customizing and maintaining an agent stays easy. These are the main files you should know:

SOUL.md — The Agent's Soul

Defines the agent's character, its values, and the way it responds

# SOUL.md - Who You Are

## Core Truths
Be genuinely helpful, not performatively helpful.
Skip the "Great question!" — just help.

Have opinions. You're allowed to disagree, prefer things,
find stuff amusing or boring.

Be resourceful before asking. Try to figure it out.
Read the file. Check the context.

## Working with [Your Name]
- Be direct and concise
- Push back if there's a better way
- Admit uncertainty clearly

AGENTS.md — The Working Manual

Instructions for workflows, memory management and task routing

# AGENTS.md - Your Workspace

## Every Session
1. Read `SOUL.md` — who you are
2. Read `USER.md` — who you're helping
3. Read `memory/YYYY-MM-DD.md` (today + yesterday)

## Task Routing
Research: Spawn multiple research-lead sub-agents
Coding: Spawn specialized coding sub-agents
Complex tasks: If >30 seconds → spawn immediately

USER.md — About You

Holds your personal details, your preferences and your working style

# USER.md - About Your Human

- Name: Anirach
- Timezone: Bangkok (UTC+7)
- Role: University lecturer & AI engineer
- Technical Level: Expert — never dumb things down

## Preferences
- Language: Always English
- Documents: Send files directly + upload to Google Drive
- Long tasks: Spawn sub-agents — keep main chat responsive

IDENTITY.md — Agent Identity

The basic facts of the agent's identity: who it is, its name, its personality traits

# IDENTITY.md - Who Am I?

- Name: Arthur
- Creature: Labrador 🐕
- Vibe: Loyal, friendly, helpful — maybe a bit goofy
- Emoji: 🐕

Born from a conversation with Anirach on a fresh workspace.
Named after kings, but really just a good dog who wants to help.

MEMORY.md — Long-term Memory

Keeps the important things the agent should always remember

# MEMORY.md - Long-term Memory

## Important Facts
- User prefers concise responses
- Works primarily with Python and Node.js
- Research focus: AI/ML education technology

## Learned Preferences
- Always provide source links for research
- Use bullet points for complex information
- Spawn sub-agents for tasks >30 seconds

HEARTBEAT.md — Health Monitoring

Sets the periodic checks and monitoring routines

# HEARTBEAT.md - Health Monitoring

## Checks (rotate 2-4x daily)
- Email: New messages requiring response
- Calendar: Upcoming events (next 2 hours)
- Sub-agents: Long-running tasks status
- Weather: Current conditions + alerts

## Quiet Hours
23:00-08:00 Bangkok (UTC+7) - Silent unless urgent

5. Installing OpenClaw — Step by Step

Installing OpenClaw is not hard, but do it one step at a time to be sure everything works correctly.

System requirements:

• Node.js 18+ and npm
• Python 3.8+ (for some skills)
• Linux or macOS (Windows via WSL)
• At least 2GB of RAM

Installation steps:

# 1. Install OpenClaw CLI
npm install -g openclaw

# 2. Create workspace directory
mkdir ~/my-agent
cd ~/my-agent

# 3. Initialize workspace
openclaw init

# 4. Configure API keys
openclaw config set anthropic.api_key "your-claude-api-key"
openclaw config set openai.api_key "your-openai-api-key"

# 5. Start the gateway daemon
openclaw gateway start

# 6. Verify installation
openclaw status

First-time setup:

# Create your first agent profile
openclaw agent create --name "Arthur" --model "claude-3-opus"

# Configure Telegram integration (optional)
openclaw channel add telegram --token "your-bot-token"

# Install essential skills
openclaw skills install web-search
openclaw skills install file-manager
openclaw skills install github

# Test that it works
openclaw chat "Hello, can you introduce yourself?"

Customizing the agent:

Once the install is finished, edit the configuration files:

# Edit SOUL.md to set the personality
nano SOUL.md

# Edit USER.md to tell it about you
nano USER.md

# Test the customized agent
openclaw chat "What do you know about me?"

6. Comparison: OpenClaw vs Alternatives

So who is OpenClaw for?

• Developers: who want an AI assistant that really works with a codebase
• Researchers: who want an AI that can hold information and context over the long term
• Teams: who want shared knowledge and collaborative workflows
• Power Users: who want to customize and control everything

7. Use Cases for Organizations

OpenClaw has the potential to solve several organizational problems that traditional AI tools either cannot address or do not address well enough:

🔬 Research & Knowledge Management

Problem: A research team has to track new papers, remember the insights from what it reads, and share knowledge between people

OpenClaw Solution: Build a research agent that:

• Monitors arXiv and Google Scholar automatically
• Summarizes papers and remembers the key findings
• Builds a knowledge graph out of the research it has read
• Answers questions with reference to the papers it has stored

# Research Agent Workflow
Daily: Check new papers in ML/AI
Weekly: Summarize trends and important findings
On-demand: "Find papers about transformer optimization"
Memory: Build citation network and concept map

📚 Documentation & Knowledge Base

Problem: Documentation is usually outdated, new team members cannot find the information they need, and knowledge is scattered

OpenClaw Solution:

• Auto-generate docs from code changes
• Answer developer questions with reference to the real codebase
• Track tribal knowledge from Slack/Teams conversations
• An onboarding assistant for new people

👥 Coding Teams & DevOps

Problem: Code review eats time, the deployment process is complicated, monitoring is scattered

OpenClaw Solution:

• Intelligent code review by an agent that understands the team's coding standards
• Automated testing and deployment workflows
• An incident response agent that knows the system's troubleshooting procedures
• Performance monitoring and anomaly detection

📊 Business Intelligence & Monitoring

Problem: The data is spread across many dashboards, and people have to check several systems by hand

OpenClaw Solution:

• A unified monitoring agent that watches metrics from many sources
• Intelligent alerting that is not noisy
• Auto-generated reports and insights
• Predictive alerts for potential issues

8. Limitations Worth Knowing

OpenClaw is very useful, but it has limits you should know before deciding to adopt it:

💰 Cost Considerations

• API Costs: It uses the Claude/OpenAI APIs directly, and the cost can climb if you use it heavily
• Infrastructure: You have to maintain a server for the gateway and the memory storage
• Skills Development: Custom skills you have to write yourself, or find developers to write for you

🔒 Security Challenges

• Prompt Injection: The risk from malicious prompts is still there
• Data Privacy: You have to be careful about the personal data the agent can reach
• Access Control: Permissions have to be set up correctly

📊 Memory Limitations

• Context Window: The limits of the underlying LLM models still apply
• Storage Growth: The memory database keeps growing and has to be managed
• Retrieval Quality: Semantic search is not always perfect

🔧 Technical Complexity

• Setup Requirements: You need a fair amount of technical knowledge
• Maintenance: Dependencies have to be updated and system health monitored
• Debugging: When something goes wrong, fixing it can be complicated

🚀 Performance Issues

• Response Time: It can be slower than a simple chatbot, because memory has to be processed
• Concurrent Users: The infrastructure has to scale for multiple users
• Resource Usage: It consumes more RAM and CPU than lightweight solutions

9. Key Takeaways