| layout | default | ||||
|---|---|---|---|---|---|
| title | Hermes Agent Tutorial | ||||
| nav_order | 42 | ||||
| has_children | true | ||||
| format_version | v2 | ||||
| source_repo | https://github.com/nousresearch/hermes-agent | ||||
| categories |
|
||||
| related_tutorials |
|
||||
| last_updated | 2026-04-12 |
NousResearch's self-hosted personal AI agent with persistent memory, autonomous skill creation, 20+ platform gateway, and a closed reinforcement-learning loop that turns every conversation into fine-tuning data.
Hermes Agent is the successor to OpenClaw — NousResearch's production-grade, self-hosted personal AI agent designed to run 24/7 on your own hardware or cloud infrastructure. With 65,972 GitHub stars and an MIT license, it represents the current state of the art in open-source agent frameworks that combine a richly layered memory system, a multi-platform messaging gateway, and a reinforcement-learning pipeline that continuously improves the underlying models through real usage.
Unlike ephemeral chatbot wrappers, Hermes is built around three design principles:
- Continuity — sessions persist, memories accumulate, skills compound. The agent you run today is smarter than the one you ran last week.
- Reach — one agent, 20+ platforms. Whether you message through Telegram, Discord, Slack, WhatsApp, Signal, Email, Matrix, Feishu, DingTalk, or a raw webhook, the same memory and skill set is available.
- Closed learning — every real interaction is a potential training example.
trajectory.pyrecords tool calls and outcomes in Atropos RL format; those trajectories can be fed directly into NousResearch's fine-tuning pipeline to improve future model behavior.
- repository:
nousresearch/hermes-agent - stars: about 225k
- GitHub release reference:
v2026.7.30(checked 2026-08-03; release metadata on GitHub)
| Audience | What You Will Get |
|---|---|
| Individual developers | A self-hosted AI assistant with memory that actually persists across sessions |
| Platform builders | A messaging gateway you can point at any of 20+ chat platforms with a single config |
| ML researchers | A live data-generation pipeline producing Atropos-format RL trajectories from real agent interactions |
| DevOps / infra engineers | Six swappable terminal backends (local, Docker, SSH, Daytona, Singularity, Modal) for isolated task execution |
| OpenClaw users | A clear migration path: hermes claw migrate imports your memories, skills, and config |
Hermes Agent is a useful study target because it combines several hard agent problems in one repo: memory, platform routing, scheduling, tool execution, and training-data capture. That makes it a strong map for readers who want to move from a local assistant to an always-on agent system.
By the end of this tutorial, you will understand how Hermes structures persistent memory, how the messaging gateway routes conversations across platforms, how scheduled jobs and subagents run, and how interaction traces become reinforcement-learning data.
Think of Hermes as a personal AI operating system. The agent core decides what to do, the memory layers preserve context, the gateway lets users reach it from many channels, and the training loop converts real usage into improvement data.
cli.py
└── hermes_cli/
├── agent/ # LLM core
│ ├── prompt_builder.py
│ ├── context_engine.py
│ ├── memory_manager.py
│ ├── skill_utils.py
│ ├── trajectory.py
│ └── smart_routing.py
├── gateway/ # 20+ platform messaging
│ ├── telegram.py
│ ├── discord.py
│ ├── slack.py
│ ├── whatsapp.py
│ ├── signal.py
│ ├── email.py
│ ├── matrix.py
│ ├── api_server.py
│ └── ...
├── cron/ # Scheduler + jobs
│ ├── scheduler.py
│ └── jobs/
├── environments/ # RL training, benchmarks, subagents
│ ├── hermes_swe_env/
│ ├── tblite/
│ └── batch_runner.py
└── acp_adapter/ # Agent Communication Protocol server
┌─────────────────────────────────────────────────────────┐
│ Memory Architecture │
├──────────────┬──────────────────┬───────────────────────┤
│ Episodic │ Semantic │ Procedural │
│ │ │ │
│ FTS5 SQLite │ MEMORY.md │ SKILL.md files │
│ session │ USER.md │ (auto-created and │
│ search + │ Honcho user │ self-improved by │
│ LLM summary │ modeling │ the agent) │
│ injection │ (dialectic) │ │
└──────────────┴──────────────────┴───────────────────────┘
| Chapter | Title | Key Topics |
|---|---|---|
| 1 | Getting Started | Install, hermes setup, ~/.hermes/ layout, first conversation, OpenClaw migration |
| 2 | The TUI and Conversation Interface | curses UI, slash commands, SOUL.md persona, context files, skin system |
| 3 | Agent Core: Prompt Building, Context Engine, Model Routing | prompt_builder.py, context_engine.py, smart_model_routing.py, credential_pool.py |
| 4 | Memory, Skills, and the Learning Loop | Three memory layers, memory_manager.py, FTS5, Honcho, SKILL.md, agentskills.io |
| 5 | The Messaging Gateway | 20+ platform drivers, session routing, delivery pipeline, API server mode |
| 6 | Cron Scheduling, Subagents, and Automation | scheduler.py, cron commands, subagent spawning, terminal backends |
| 7 | RL Training and Trajectory Generation | trajectory.py, Atropos, benchmark envs, tool-call parsers, data pipeline |
| 8 | ACP, MCP, Migration, and Ecosystem | ACP server, MCP integration, agentskills.io, OpenClaw migration, Nix/Docker deploy |
# Install
curl -fsSL https://raw.githubusercontent.com/nousresearch/hermes-agent/main/install.sh | bash
# Run setup wizard
hermes setup
# Start the TUI
hermes| Feature | Hermes Agent | LangChain | AutoGPT | CrewAI |
|---|---|---|---|---|
| Persistent episodic memory (FTS5) | Yes | Plugin-dependent | Partial | No |
| Autonomous skill creation | Yes | No | No | No |
| 20+ platform gateway | Yes | No | No | No |
| RL trajectory generation | Yes | No | No | No |
| Closed fine-tuning loop | Yes | No | No | No |
| Self-hosted, MIT license | Yes | Yes | AGPL | MIT |
| Six terminal backends | Yes | No | No | No |
| ACP multi-agent protocol | Yes | No | No | No |
Hermes Agent is released under the MIT License by NousResearch. This tutorial is an independent educational resource; it is not officially affiliated with NousResearch.