Hermes Agent + Honcho: Theory-of-Mind Memory for Hermes

What Is Honcho?

Honcho from Plastic Labs is a personalization memory layer for AI agents. Instead of just summarizing past messages, Honcho builds a representation of each user — preferences, beliefs, communication style — and updates it as the agent learns more. Open-source, self-hostable, and provider-agnostic.

Why Combine It With Hermes?

Hermes Agent has strong workspace memory: file system, embedding search, conversation history. What Hermes doesn't do natively is build a theory of mind — an evolving model of who each user is and what they want. Honcho fills that gap.

The combination:

• Hermes handles the chat surface, tool calls, file work, and shell access.
• Honcho persists user-level personality, preferences, and history across sessions.
• Hermes calls Honcho via its HONCHO_API_KEY + HONCHO_APP_ID environment variables (or via a custom skill / MCP server).

Option 1: Hermes on OpenClaw Launch (Easiest)

Hermes Hosting on OpenClaw Launch gives you a managed Hermes container with $HOME/.hermes persistence built in. To wire Honcho:

1. Deploy a Hermes instance from openclawlaunch.com/hermes-hosting.
2. Open your container shell from the dashboard.
3. Set Honcho env vars in ~/.hermes/.env:

HONCHO_API_KEY=ph_...
HONCHO_APP_ID=hermes-bot
HONCHO_BASE_URL=https://demo.honcho.dev

4. Add a small skill or MCP tool that calls Honcho on each user message:

# pseudocode inside a Hermes tool/skill
context = honcho.get_user_context(user_id)
hermes.system_prompt += f"\nUser context: {context}"
honcho.observe(user_id, latest_message)

Option 2: Self-Hosted Hermes

export HONCHO_API_KEY=ph_...
export HONCHO_APP_ID=hermes-bot

# In hermes config.yaml
tools:
  mcp:
    - name: honcho
      command: npx
      args: ["-y", "@honcho-ai/mcp-server"]

The Honcho MCP server exposes get_user_representation and observe_message tools that Hermes can call automatically once they're registered.

Workflow: How a Personalized Hermes Conversation Looks

1. User sends a message via Telegram/Discord/web chat into Hermes.
2. Hermes's pre-prompt hook calls honcho.get_user_representation(user_id).
3. Hermes injects the user's persona/preferences into the system prompt.
4. Hermes reasons + replies with the persona context applied.
5. Hermes's post-prompt hook calls honcho.observe(user_id, exchange).
6. Honcho updates its theory of mind for that user.

When to Use Honcho With Hermes

• You run a Hermes bot for multiple users (community, customer support, team)
• You want the agent to remember preferences across days/weeks
• You need personalization beyond raw conversation history (preferences, beliefs, tone)
• You're building a long-running personal assistant that should "know" the user better over time

Hermes's Built-In Memory vs Honcho

What's Next?

• Hermes Agent Memory — built-in memory features and how they work
• Hermes Agent + MCP — how to wire MCP servers into Hermes
• OpenClaw vs Honcho — when to pick which
• Hermes Agent Skills — build a custom skill that calls Honcho directly