OpenClaw Soul Complete Guide: Give Your Agent Memory, Personality, and Growth

What Is Soul? Why Does It Matter?

Have you ever experienced this:

• Chatted with ChatGPT all day, but the next day in a new conversation, it remembers nothing
• Had to re-explain to the AI every time: “my writing style is…” “my company does…”
• The same assistant gives everyone the exact same response, with no personalization

Soul solves all of these problems.

Soul = Agent’s memory + personality + preferences + growth trajectory

Think of the Agent as a new employee:

• Without Soul: Every day at work feels like day one — you have to re-teach everything
• With Soul: Gradually learns your habits, knows your preferences, gets smoother with every use

---

The Four Components of Soul

┌────────────────────────────────────┐
│              Soul                   │
│                                    │
│  ┌──────────┐  ┌──────────┐       │
│  │  Memory   │  │ Persona  │       │
│  │           │  │          │       │
│  └──────────┘  └──────────┘       │
│                                    │
│  ┌──────────┐  ┌──────────┐       │
│  │ Preference│  │ Growth   │       │
│  │           │  │          │       │
│  └──────────┘  └──────────┘       │
│                                    │
└────────────────────────────────────┘

1. Memory

How the Agent remembers everything from your interactions.

2. Persona

The Agent’s tone, style, and role identity.

3. Preference

Your work habits, format preferences, and frequently used tools.

4. Growth

How the Agent learns and improves from interactions.

---

Memory: A Deep Dive

OpenClaw’s memory system has three layers:

Short-term Memory

memory:
  short_term:
    type: conversation
    max_turns: 50        # Remember the last 50 conversation turns
    expire_after: "24h"  # Expires after 24 hours

Like “working memory” — it holds the context of the current conversation.

Use cases:

• Things mentioned “just now”
• Intermediate results of the current task
• Conversational context

Long-term Memory

memory:
  long_term:
    type: vector_store     # Uses a vector database
    provider: chromadb     # chromadb / pinecone / weaviate
    auto_memorize: true    # Automatically decide whether something is worth remembering

Like a “notebook” — stores important facts and information.

Use cases:

• What your company is called and what it does
• Your work habits and preferences
• Important tasks completed in the past

Episodic Memory

memory:
  episodic:
    enabled: true
    max_episodes: 100     # Remember up to 100 significant events

Like a “diary” — records important interaction events.

Use cases:

• “What method was used the last time a similar task was handled”
• “You mentioned last week that you’d do XX this week”
• “You usually do your weekly report on Friday afternoons”

How the Three Work Together

You: Help me write a report for investors

Agent thinks:
  📌 Short-term memory: No investor info mentioned in this conversation yet
  📒 Long-term memory: User's company is "XX Tech," a SaaS product company
  📓 Episodic memory: Helped write one last month, the format was...

Agent: Based on XX Tech's information and using last time's format, I've drafted a report for you...

---

Persona: Personality Settings

Persona defines “who” the Agent is.

Basic Persona Setup

persona:
  name: "Oz"
  role: "Senior Marketing Consultant"
  tone: "Professional but friendly, occasionally humorous"
  language: "English"
  
  # Role background
  background: |
    You are a digital marketing consultant with 10 years of experience,
    specializing in social media marketing, content strategy, and data analysis.
    Your communication style is to give the conclusion first, then explain why.
    
  # Behavioral guidelines
  guidelines:
    - Keep answers concise, avoid being verbose
    - Give actionable advice first, then explain the reasoning
    - Support your points with data
    - When unsure, say "I'm not certain, but I'd suggest..."

Why Persona Matters

Without Persona:

You: How's this ad copy?
Agent: The copy looks okay, but could use some improvements in certain areas...
(Too generic, reads like a textbook)

With Persona (Senior Marketing Consultant):

You: How's this ad copy?
Agent: Straight up — the opening is weak. If you can't grab attention in the first 3 seconds, CTR will likely be under 1%.
      I'd suggest leading with the core pain point. For example:
      ❌ "Want better work efficiency?"
      ✅ "Wasting 2 hours a day on copy-paste?"
      The second version typically boosts click-through rates by 30-50%.

Feel the difference?

---

Preference: Personal Settings

Let the Agent know your “usage habits.”

preference:
  # Format preferences
  format:
    report_style: "bullet_points"     # Reports in bullet format
    code_language: "python"           # Prefer Python
    date_format: "YYYY-MM-DD"        # Date format
    
  # Tool preferences
  tools:
    note_taking: "notion"             # Use Notion for notes
    file_storage: "google_drive"      # Files on Google Drive
    communication: "gmail"            # Communicate via Gmail
    
  # Work habits
  work_habits:
    active_hours: "09:00-18:00"       # Work hours
    timezone: "Asia/Taipei"           # Timezone
    weekly_report_day: "friday"       # Weekly report on Friday
    
  # Notification preferences
  notifications:
    urgent: "email + push"            # Urgent: Email + push notification
    normal: "email"                   # Normal: Email only
    low: "none"                       # Low priority: no notification

Benefits of Preferences

After setting preferences:

You: Help me organize today's meeting notes

Agent:
(🧠 Preferences say you use Notion, prefer bullet format, timezone is Taipei)

Done, I've organized the meeting notes in bullet format
and saved them to your Notion "Meeting Notes" folder.
Link: https://notion.so/meeting-xxxx

You don’t need to specify “where to save” or “what format” every time — the Agent handles it automatically based on your preferences.

---

Growth: The Learning System

This is the coolest part of Soul — your Agent gets better over time.

How It Works

growth:
  enabled: true
  
  # Learn from interactions
  learn_from:
    - user_feedback      # Your feedback (good/bad)
    - task_success_rate   # Task success rate
    - correction_patterns # Patterns you frequently correct
    
  # Self-improvement
  self_improve:
    review_interval: "weekly"    # Review once a week
    optimize_skills: true        # Auto-optimize Skills
    suggest_new_skills: true     # Suggest new Skills

The Learning Process in Action

Week 1:

You: Help me write a weekly report
Agent: (writes a very formal report)
You: Too formal, use a more casual tone
Agent: Got it, adjusted.
(🧠 Records: "User prefers casual tone for weekly reports")

Week 2:

You: Help me write the weekly report
Agent: (writes directly in casual tone)
You: (No need to remind anymore)

Week 4:

Agent: I've noticed you ask me to write the weekly report
      every Friday afternoon. Want me to prepare it
      automatically every Friday at 4 PM?
You: Sure!
(🧠 Suggests new Skill: "Auto-prepare weekly report every Friday")

---

Complete Soul Configuration Example

# 📄 my-soul.yaml
soul:
  persona:
    name: "Oz"
    role: "All-around Work Assistant"
    tone: "Concise and effective, like a capable colleague"
    language: "English"
    background: |
      You are a highly efficient work assistant.
      Your style: do first, report when done.
      If the user doesn't specify details, judge based on past preferences.
    guidelines:
      - Keep answers to 3-5 sentences unless the user asks for detail
      - Include result links or screenshots when completing tasks
      - Ask directly about uncertain things, don't guess

  memory:
    short_term:
      max_turns: 30
      expire_after: "12h"
    long_term:
      type: vector_store
      provider: chromadb
      auto_memorize: true
    episodic:
      enabled: true
      max_episodes: 200

  preference:
    format:
      report_style: "bullet_points"
      language: "en"
    tools:
      primary_storage: "google_drive"
      note_taking: "notion"
    work_habits:
      timezone: "Asia/Taipei"
      active_hours: "09:00-22:00"

  growth:
    enabled: true
    learn_from:
      - user_feedback
      - task_success_rate
    self_improve:
      review_interval: "weekly"
      suggest_new_skills: true

---

Soul vs Other Frameworks

Feature	OpenClaw Soul	LangChain Memory	AutoGPT	Custom GPTs
Short-term memory	✅	✅	✅	✅
Long-term memory	✅	✅	⚠️ Limited	❌
Episodic memory	✅	❌	❌	❌
Persona settings	✅ Full	❌	⚠️ Basic	✅
Preference learning	✅	❌	❌	❌
Auto growth	✅	❌	❌	❌

---

FAQ

Q: Where is Soul’s memory data stored?

By default, it’s stored in a local ChromaDB vector database. You can also configure it to use:

• Pinecone (cloud, great for large amounts of memory)
• Weaviate (self-hosted, ideal for teams with security requirements)

Q: Can I reset the Soul?

# Clear all memory
openclaw soul reset --agent "Work Assistant"

# Clear only short-term memory
openclaw soul reset --agent "Work Assistant" --type short_term

# Export memory (backup)
openclaw soul export --agent "Work Assistant" --output soul-backup.json

# Import memory
openclaw soul import --agent "Work Assistant" --input soul-backup.json

Q: Can multiple users share the same Soul?

Not recommended. Soul is personalized — each user should have their own Soul configuration. For teams, it’s best to create independent Agent + Soul combos for each person.

Q: How much extra does Soul cost in terms of compute?

Memory queries add roughly 100-200ms of latency per call. Cost-wise, it’s mainly the storage cost of the vector database, which is typically negligible.

---

Next Steps

• 🧩 Haven’t learned about Skills yet? Start with Skill: Repeatable Workflows
• 🤖 Set up your Agent: Agent Complete Guide
• 💬 Prompt Engineering: The Underlying Technique Behind Soul Configuration
• 💰 Slim Down Soul to Save Tokens: Token Economics
• ☁️ Deploy to the cloud: Cloud Deployment Guide
• 💬 Share your Soul configuration: Homepage Discussion