OpenClaw Integration Hub

Three battle-tested integration packages that make OpenClaw agents smarter: MemPalace for persistent memory, Hermes P1-P5 for self-improvement, and Claw-Code for permission-aware execution.

Why This Exists

These integrations were built to solve real problems running multi-agent AI systems. After months of production use with agents Lisa, Nyx, and Kael, these patterns proved their value — Lisa went from hallucinating facts to checking memory before responding, and agent execution became auditable and safe.

Integrations

Integration	What It Does	Impact
MemPalace	Persistent AI memory with semantic search	96.6% LongMemEval R@5, 27K+ drawers, agents stop hallucinating
Hermes P1-P5	Five self-improvement systems for agents	Auto-skill creation, memory nudges, dynamic user modeling
Claw-Code	Permission-aware tool execution + cost tracking	3-tier permissions, per-agent cost tracking, 207 commands

Quick Start

1. MemPalace MCP Server

pip install mempalace fastmcp

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "mempalace": {
      "command": "python3",
      "args": ["/path/to/mempalace/mempalace_mcp_server.py"],
      "env": { "MEM_PALACE": "/path/to/your/palace" }
    }
  }
}

2. Hermes Self-Improvement

cp -r self-improving-agent/ ~/.openclaw/workspace/skills/
mkdir -p ~/.openclaw/workspace/.learnings

3. Claw-Code Permissions

cd claw-code/
python3 bootstrap.py

Architecture

openclaw-integration/
├── mempalace/               # MemPalace MCP integration
│   ├── mempalace_mcp_server.py  # FastMCP server (19 tools)
│   ├── mempalace_cli.py     # CLI wrapper for non-MCP agents
│   └── SKILL.md             # Usage documentation
├── hermes/                  # Hermes P1-P5 systems
│   ├── p1-memory-nudge/     # Periodic memory reminders
│   ├── p2-auto-skill-creation/ # Auto-extract skills from patterns
│   ├── p3-dynamic-user-model/ # Adaptive user profiling
│   ├── p4-fts5-search/      # Fast SQLite full-text search
│   ├── p5-skill-quality/    # Skill quality tracking & scoring
│   └── lisa_bridge_monitor.py # Real-time Claude <-> Lisa bridge
├── claw-code/               # Claw-Code permission integration
│   ├── claw_integration.py  # Core integration module
│   ├── openclaw_bridge.py   # Agent bridge with hooks
│   ├── bootstrap.py         # Initialization script
│   └── skill.json           # Metadata
├── self-improving-agent/    # Full self-improvement skill
│   ├── SKILL.md             # Complete skill documentation
│   ├── assets/              # Templates
│   ├── hooks/               # Hook scripts
│   ├── references/          # Setup guides
│   └── scripts/             # Automation scripts
└── README.md

Results in Production

After deploying these integrations to the OpenClaw multi-agent system:

• Lisa (Authority agent): Memory lookups before responding eliminated hallucinated facts
• Nyx (Intelligence agent): Self-improvement logging catches recurring issues
• Kael (Execution agent): Permission-aware execution prevents accidental damage
• Cost visibility: Per-agent token tracking revealed that 60% of spend was on redundant queries — now optimized

Requirements

• Python 3.10+
• mempalace package (pip install mempalace)
• fastmcp package (pip install fastmcp)
• ChromaDB (auto-installed with mempalace)

License

MIT

Credits

Built for the OpenClaw multi-agent system. Self-improving agent skill based on pskoett/pskoett-ai-skills.