Summary

This repository enables self-hosting of Honcho, the memory layer for Hermes Agent, providing data sovereignty and control over LLM providers. It targets Hermes users seeking enhanced privacy or custom LLM backends, offering a balance between data security and AI capability by integrating with any OpenAI-compatible API or local LLM inference.

How It Works

Honcho's stack (API, Deriver, PostgreSQL, Redis) is deployed locally via Docker. Components like Deriver (observation extraction), Dialectic (recall), Summary (context compression), and Dream (consolidation) utilize LLMs. LLM calls are routed through user-configured endpoints or local servers, ensuring user data remains on-premises.

Quick Start & Requirements

Installation is streamlined via a setup script: curl -sL https://raw.githubusercontent.com/elkimek/honcho-self-hosted/main/setup.sh -o /tmp/setup.sh && bash /tmp/setup.sh. Prerequisites include Ubuntu 22.04+, Docker, and an OpenAI-compatible API key. Setup is estimated at ~3 minutes. Manual setup involves Docker installation, repo cloning, and API key configuration.

Highlighted Details

• LLM Flexibility: Supports any OpenAI-compatible API or local LLM inference (Ollama, vLLM), with tiered LLM usage for cost/performance optimization.
• Data Sovereignty: All user data resides locally, offering enhanced privacy compared to the default managed cloud.
• Neuromancer Context: While the default cloud uses specialized Neuromancer models, this repo allows leveraging alternative LLMs for memory tasks.
• MCP Server: An optional component exposes memory tools to clients like Claude Code/Desktop for local integration.

Maintenance & Community

Update instructions via Docker Compose and Git are provided. The README does not detail community channels or notable contributors.

Licensing & Compatibility

Licensed under GPL-3.0, a strong copyleft license that may impose restrictions on commercial use or integration with proprietary software.

Limitations & Caveats

Embedding configuration can unintentionally fall back to backup provider credentials. Only one backup provider is supported per component. End-to-end encryption is not feasible for function calling, meaning LLM providers see request content. Local inference requires 32B+ models for reliable function calling; embedding models typically need a separate cloud API.