openclaw/mem0

Mem0 Memory Integration

Mem0 adds an intelligent, adaptive memory layer to Clawdbot that automatically learns and recalls user preferences, patterns, and context across all interactions.

Core Workflow

1. Search Before Responding

Before answering user questions, search mem0 for relevant context:

node scripts/mem0-search.js "user preferences" --limit=3

Use retrieved memories to:

• Personalize responses
• Remember preferences
• Recall past patterns
• Adapt communication style

2. Store After Interactions

Explicit Storage (when user says "remember this"):

node scripts/mem0-add.js "Abhay prefers concise updates"

Conversation Storage (for context learning):

# Pass messages as JSON
node scripts/mem0-add.js --messages='[{"role":"user","content":"I like brief updates"},{"role":"assistant","content":"Got it!"}]'

Available Commands

Search Memories

node scripts/mem0-search.js "query text" [--limit=3] [--user=abhay]

Searches semantically across stored memories. Returns relevant memories ranked by relevance.

Add Memory

# Simple text
node scripts/mem0-add.js "memory text" [--user=abhay]

# Conversation messages (auto-extracts memories)
node scripts/mem0-add.js --messages='[{...}]' [--user=abhay]

Mem0's LLM automatically extracts, deduplicates, and merges related memories.

List All Memories

node scripts/mem0-list.js [--user=abhay]

Shows all stored memories for the user with IDs and creation dates.

Delete Memories

# Delete specific memory
node scripts/mem0-delete.js <memory_id>

# Delete all memories for user
node scripts/mem0-delete.js --all --user=abhay

What to Store vs Not Store

✅ Store These:

• Explicit requests: "Remember that I..."
• Preferences: Communication style, format choices
• Personal context: Work info, interests, family (non-sensitive)
• Usage patterns: Frequent requests, timing preferences
• Corrections: When user corrects your mistakes
• Adaptive facts: Current projects, recent interests

❌ Don't Store:

• Secrets, passwords, API keys
• Temporary context (unless explicitly requested)
• System errors or debug info
• Information already in MEMORY.md (avoid duplication)

Complementing Clawdbot Memory

Clawdbot MEMORY.md (Structured, Deliberate):

• Permanent facts: Name = Abhay, Location = Singapore
• Reference data: Email, blog URL, Twitter handle
• Structured knowledge: Project details, credentials

Mem0 (Dynamic, Learned):

• Preferences: "Abhay prefers concise updates"
• Patterns: "Usually asks for bus info at 8:30am"
• Adaptive context: "Currently interested in AI news"
• Behavioral: "Likes direct answers, minimal fluff"

Use both together: Check MEMORY.md for facts, check mem0 for preferences/patterns.

Performance Benefits

• +26% accuracy over OpenAI Memory (LOCOMO benchmark)
• 91% faster than full-context retrieval
• 90% fewer tokens than including all conversation history
• Sub-50ms semantic search retrieval

Configuration

Located in scripts/mem0-config.js:

{
  embedder: "openai/text-embedding-3-small",
  llm: "openai/gpt-4o-mini",
  vectorStore: "memory" (local),
  historyDb: "~/.mem0/history.db",
  userId: "abhay"
}

Uses Clawdbot's OpenAI API key from environment (OPENAI_API_KEY).

Integration Patterns

For detailed workflow patterns, error handling, and best practices, see:

• references/integration-patterns.md

Programmatic Use

All scripts support JSON_OUTPUT environment variable for programmatic access:

JSON_OUTPUT=1 node scripts/mem0-search.js "query"

Returns JSON after human-readable output (look for ---JSON--- marker).

Resources

scripts/

• mem0-config.js - Configuration and instance initialization
• mem0-search.js - Search memories semantically
• mem0-add.js - Add new memories
• mem0-list.js - List all memories
• mem0-delete.js - Delete memories

references/

• integration-patterns.md - Detailed best practices and patterns