memU Proactive Memory for DEXBot2

memU provides 24/7 always-on proactive memory for AI agents integrated with DEXBot2. It captures user intent, reduces LLM token costs, and enables context-aware trading assistance.

Overview

memU treats memory like a file system — structured, hierarchical, and instantly accessible:

Memory Hierarchy

memory/
├── preferences/
│   ├── communication_style.md
│   └── trading_preferences.md
├── relationships/
│   ├── contacts/
│   └── interaction_history/
├── knowledge/
│   ├── domain_expertise/
│   └── trading_strategies/
└── context/
    ├── recent_conversations/
    └── pending_tasks/

Prerequisites

• Python 3.13+
• memU package: pip install memu-py
• LLM API key (OpenAI, OpenRouter, etc.)

Available Tools

Core Memory Operations

Trading-Specific Operations

Modalities

Usage Examples

Memorize a Conversation

{
  "tool": "memu_memorize_conversation",
  "arguments": {
    "messages": [
      {"role": "user", "content": "I prefer BTS/USD grid bots with 2% increment"},
      {"role": "assistant", "content": "I'll configure a grid bot with those settings"}
    ],
    "user": {"user_id": "trader-123"}
  }
}

Memorize Trading Context

{
  "tool": "memu_memorize_trading_context",
  "arguments": {
    "context": {
      "bot": "BTS/USD-grid",
      "event": "price_dropped_5_percent",
      "action_taken": "rebalanced_grid",
      "timestamp": "2026-05-18T10:30:00Z"
    },
    "user": {"user_id": "trader-123"}
  }
}

Retrieve Trading Context

{
  "tool": "memu_retrieve_trading_context",
  "arguments": {
    "query": "What are my preferences for BTS/USD grid bots?",
    "user": {"user_id": "trader-123"}
  }
}

Retrieve with LLM Deep Reasoning

{
  "tool": "memu_retrieve",
  "arguments": {
    "queries": [
      {"role": "user", "content": {"text": "How should I adjust my grid based on recent market behavior?"}}
    ],
    "method": "llm",
    "where": {"user_id": "trader-123"}
  }
}

Integration with DEXBot2 Claw

The memU bridge integrates with the DEXBot2 claw subsystem:

CLI Usage

# From claw/ directory
npm run memu:status
npm run memu:mcp  # Start MCP server

MCP Server Configuration

For Hermes:

mcp_servers:
  memu:
    command: "node"
    args: ["/path/to/DEXBot2/dist/claw/scripts/memu_mcp_server.js", "--memu-dir", "/path/to/claw/data/memu"]

For NanoBot/PicoClaw:

node ../dist/claw/scripts/memu_mcp_server.js --memu-dir /path/to/claw/data/memu

Proactive Memory Patterns

Pattern 1: Learning User Preferences

When the user mentions trading preferences:

1. Extract the preference from the conversation
2. Call memu_memorize_conversation to store it
3. On future interactions, call memu_retrieve_trading_context to recall preferences

Pattern 2: Trading Event Memory

When significant trading events occur:

1. Call memu_memorize_trading_context with event details
2. Store bot actions, market conditions, and outcomes
3. Later, retrieve to inform similar decisions

Pattern 3: Context-Aware Assistance

Before responding to trading queries:

1. Call memu_retrieve_trading_context with the query
2. Use retrieved context to provide personalized responses
3. Reference past decisions and preferences

Memory Types

Best Practices

1. Scope memories to users: Always pass user with user_id when multiple users share the system
2. Use appropriate modalities: Choose the right modality for the content type
3. Retrieve before acting: Check existing memories before making recommendations
4. Use RAG for speed, LLM for depth: method: "rag" is fast, method: "llm" provides deeper reasoning
5. Clean up old memories carefully: Prefer scoped memu_clear calls such as {"where":{"user_id":"trader-123"}} on shared systems

Environment Variables

Notes

• memU runs as a Python subprocess bridge from Node.js
• Memory is persisted to SQLite under claw/data/memu by default
• For production, configure PostgreSQL with pgvector for persistent storage
• The MCP server uses stdio transport with newline-delimited JSON-RPC