zengyi-thinking M2Wise - Memory-to-Wisdom Engine
Enable AI companion robots to grow wisdom from experience
What This Skill Does
M2Wise extracts memories from conversations and generates actionable wisdom that evolves over time. It's designed for AI companions that need persistent user understanding.
When to Use
Use M2Wise when you need:
- Long-term user memory (preferences, facts, commitments)
- Personalized wisdom generation from accumulated interactions
- Self-improving AI behavior based on feedback
- Context-aware responses that remember user patterns
Quick Reference
| Situation | Action |
|---|---|
| User expresses preference | engine.add() → memory extracted as preference type |
| User provides factual info | engine.add() → memory extracted as fact type |
| Need contextual memory | engine.search(query, user_id) → returns memory + wisdom |
| Generate wisdom | engine.sleep(user_id) → creates wisdom drafts |
| Publish wisdom | engine.dream(user_id) → verifies and publishes |
| Check effectiveness | Automatic confidence tracking via Dream phase |
Installation
Prerequisites
# Python 3.11+
python3 --version
# Install M2Wise
pip install m2wise
# Or install with all features
pip install m2wise[all]
Environment Setup
# Option 1: OpenAI
export OPENAI_API_KEY="sk-..."
# Option 2: SiliconFlow (recommended for Chinese)
export M2WISE_SILICONFLOW_API_KEY="sk-..."
# Option 3: Anthropic
export ANTHROPIC_API_KEY="sk-..."
OpenClaw Integration
# Manual installation
git clone https://github.com/your-repo/m2wise.git ~/.openclaw/skills/m2wise
# Or via ClawHub
npx clawdhub@latest install m2wise
Core Usage
Basic Workflow
from m2wise import M2Wise, M2WiseConfig
# Initialize
config = M2WiseConfig(data_dir="./m2wise_data")
engine = M2Wise(config=config)
# 1. Add conversation
engine.add(
[{"role": "user", "content": "I prefer concise Chinese answers for technical questions"}],
user_id="alice"
)
# 2. Search memories and wisdom
bundle = engine.search("How to debug Python?", user_id="alice")
print(bundle.as_prompt())
# 3. Generate wisdom (Sleep phase)
sleep_report = engine.sleep(user_id="alice")
print(f"Created {sleep_report.drafts_created} wisdom drafts")
# 4. Publish wisdom (Dream phase)
dream_report = engine.dream(user_id="alice")
print(f"Published {dream_report.published} wisdoms")
SDK Usage (Recommended)
from m2wise_sdk import M2WiseSDK
sdk = M2WiseSDK()
# Add messages
sdk.add_message("alice", "I'm a Python developer")
sdk.add_message("alice", "I prefer short responses")
# Get context
context = sdk.get_context("alice", "How should I answer?")
# Trigger wisdom generation
sdk.trigger_sleep("alice")
sdk.trigger_dream("alice")
# Get stats
stats = sdk.get_stats("alice")
Memory Types
| Type | Description | Example |
|---|---|---|
preference |
User's explicit preferences | "I like concise answers" |
fact |
Factual information | "User is a software engineer" |
explicit |
Direct memory requests | "Remember that I hate spam" |
commitment |
User commitments | "I will exercise every morning" |
Wisdom Types
| Type | Description | Example |
|---|---|---|
principle |
Interaction principles | "Prefer concise Chinese answers" |
schema |
Behavioral patterns | "Technical questions need examples" |
skill |
Operational skills | "Use tmux for long-running tasks" |
causal_hypothesis |
Causal assumptions | "User delays = task is low priority" |
Three-Phase Evolution
┌────────────────────────────────────────────────────────────┐
│ Online → Sleep → Dream │
│ ↓ ↓ ↓ │
│ Add/Search Extract Verify & Publish │
│ ↓ ↓ ↓ │
│ Real-time Memory Counterexample │
│ Interaction Clustering Mining │
└────────────────────────────────────────────────────────────┘
Phase 1: Online (real-time)
- engine.add() - Store conversation
- engine.search() - Retrieve context
Phase 2: Sleep (consolidation)
- engine.sleep() - Extract memories
- Generate wisdom drafts from clusters
Phase 3: Dream (verification)
- engine.dream() - Mine counterexamples
- Verify and publish wisdom
- Auto-evolve based on hit rates
Configuration
from m2wise import M2WiseConfig
config = M2WiseConfig(
data_dir="./data", # Storage path
embedder="siliconflow", # openai, siliconflow, anthropic, local
embedder_model="BAAI/bge-large-zh-v1.5",
vector_store="faiss", # faiss, qdrant, postgres
similarity_threshold=0.7, # Retrieval threshold
max_memories=100, # Max memories to retrieve
auto_sleep=True, # Auto-trigger sleep
auto_dream=True, # Auto-trigger dream
cache_enabled=True, # Enable caching
cache_ttl=3600, # Cache TTL (seconds)
)
MCP Server Tools
When using MCP server:
| Tool | Description |
|---|---|
m2wise_add |
Add memory from conversation |
m2wise_search |
Search memories and wisdom |
m2wise_sleep |
Generate wisdom drafts |
m2wise_dream |
Verify and publish wisdom |
m2wise_list_wisdom |
List all wisdom |
m2wise_forget |
Delete memory or wisdom |
m2wise_stats |
Get user statistics |
m2wise_chat |
Full conversation interaction |
Adapter Support
| Adapter | Status | Description |
|---|---|---|
| M2Wise | ✅ Native | Native format |
| mem0 | ✅ Compatible | mem0.ai compatible |
| Letta | ✅ Compatible | Letta memory format |
| Anthropic | ✅ Compatible | Anthropic messages |
Detection Triggers
Automatically detect these patterns:
Preferences:
- "I prefer..."
- "I like..."
- "Don't use..."
- "Please use..."
Facts:
- "I'm a..."
- "I work as..."
- "I have..."
Commitments:
- "I will..."
- "I promise..."
- "I'm going to..."
Best Practices
- Regular Sleep: Call
engine.sleep()periodically to extract memories - Dream Verification: Call
engine.dream()to verify and publish wisdom - Confidence Tracking: Wisdom confidence auto-evolves based on hit rates
- Use SDK: Use M2WiseSDK for simpler API
Performance
| Component | Performance |
|---|---|
| Similarity Calculation | ~300x speedup with caching |
| Confidence Evaluation | 1000 evaluations/10ms |
| Full Workflow | 45000+ memories/sec |
Related Resources
- API Reference: See REFERENCE.md
- Examples: See EXAMPLES.md
- GitHub: https://github.com/your-repo/m2wise