Hermes Agent Official Docs Explained vs OpenClaw
Important: This post centers on the official documentation, replacing the earlier version based solely on video descriptions. Source priority: official docs > authoritative analysis > video descriptions.
1. Product Positioning
Hermes Agent is an open-source AI Agent developed by Nous Research (the Hermes-3 / Nomos / Psyche model family), under the MIT license, released in February 2026.
Website: https://hermes-agent.nousresearch.com/
GitHub: roughly 22,000 stars
Core philosophy:
“It’s not a coding copilot tethered to an IDE or a chatbot wrapper around a single API. It’s an autonomous agent that gets more capable the longer it runs.”
Key positioning differences from OpenClaw:
- OpenClaw = a local-first Agent orchestration framework, strong in multi-channel integration, team collaboration, and enterprise governance
- Hermes Agent = a self-evolving personal operator, strong in long-term memory, automatic skill acquisition, and user-preference modeling
2. Core Architecture (Official)
Code scale
| Component | File | Scale |
|---|---|---|
| AIAgent (conversation loop) | run_agent.py | ~9,200 lines |
| HermesCLI (interactive terminal) | cli.py | ~8,500 lines |
| Gateway (message gateway) | gateway/run.py | ~5,800 lines |
| Config commands | hermes_cli/main.py | ~4,200 lines |
| Interactive install wizard | hermes_cli/setup.py | ~3,500 lines |
| Test suite | tests/ | 3,000+ tests |
6 terminal backends
| Backend | Purpose |
|---|---|
| local | Execute directly on the local machine |
| Docker | Containerized, isolated execution |
| SSH | Remote server |
| Daytona | Serverless persistence |
| Modal | Serverless persistence |
| Singularity | HPC containers |
Serverless behavior: Daytona and Modal support hibernation — near-zero cost when the environment is idle.
3. Memory System (Official Details)
Two-layer memory files
| File | Purpose | Capacity |
|---|---|---|
MEMORY.md |
The agent’s personal notes: environment facts, conventions, lessons learned | 2,200 chars (~800 tokens) |
USER.md |
User profile: preferences, communication style, expectations | 1,375 chars (~500 tokens) |
Skill system (Self-Improving Skills)
Skills are knowledge documents loaded on demand, following a progressive disclosure pattern. The agent automatically creates skills at these moments:
- After completing a complex task (5+ tool calls)
- After hitting an error or dead end and finding a workable path
- After the user corrects the agent’s approach
- When it discovers a non-trivial workflow
4. MCP (Model Context Protocol) Integration
Two kinds of MCP servers supported
| Type | Configuration | Use case |
|---|---|---|
| Stdio (local subprocess) | command + args + env |
Locally installed, low latency |
| HTTP (remote endpoint) | url + headers |
Hosted services, organization-internal MCP |
Dynamic tool discovery
An MCP server can notify Hermes of changes to its tool list via notifications/tools/list_changed; Hermes automatically re-fetches and updates its registry, with no manual reload needed.
5. Voice Mode
Three voice features
| Feature | Platform | Description |
|---|---|---|
| Interactive Voice | CLI | Ctrl+B to record, voice interruption, streaming TTS |
| Auto Voice Reply | Telegram, Discord | Sends voice audio alongside text replies |
| Voice Channel | Discord | Joins a VC, listens to users speaking, and replies with voice |
Local STT (zero API cost)
pip install faster-whisper # 免费，运行本地，约 150MB 模型，首次使用自动下载
6. The SOUL.md Persona System
SOUL.md vs AGENTS.md
| Purpose | SOUL.md | AGENTS.md |
|---|---|---|
| Identity/persona | ✅ | ❌ |
| Tone/style | ✅ | ❌ |
| Communication preferences | ✅ | ❌ |
| Project architecture | ❌ | ✅ |
| Code conventions | ❌ | ✅ |
| Tool preferences | ❌ | ✅ |
| Repo-specific workflows | ❌ | ✅ |
The rule of thumb: if it should follow you everywhere → SOUL.md; if it belongs to a specific project → AGENTS.md
7. 14+ Messaging Platforms
Officially supported platforms (one Gateway, a unified experience):
CLI, Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Mattermost, Email, SMS, DingTalk, Feishu, WeCom, Home Assistant
Signal and Feishu are Hermes’ unique differentiating channels.
8. Full Comparison with OpenClaw
Feature comparison table
| Feature | Hermes Agent | OpenClaw |
|---|---|---|
| Memory system | Multi-layer (Active + Archive + Honcho), FTS5 retrieval, ~3,300 chars across two files | Isolated memory per Assistant |
| Skill system | 47 tools, agent automatically creates/improves skills | 52+ built-in Skills, file precedence |
| Self-evolution | ✅ Automatically writes a Skill file when a task completes | ❌ Does not auto-generate new skills |
| User modeling | Honcho dialectic user modeling | ❌ |
| Deployment backends | 6 (local/Docker/SSH/Daytona/Modal/Singularity) | Managed + cloud containers |
| Serverless hibernation | ✅ Zero cost when Daytona/Modal are idle | ❌ |
| Channel count | 14+ (Signal + Feishu unique) | Mainstream messaging platforms |
| Model support | 200+ via OpenRouter/Nous Portal/Ollama | BYOK: Claude/GPT/Gemini/xAI/Groq/Mistral |
| MCP support | ✅ Native MCP, stdio + HTTP, dynamic discovery | ❌ (not native) |
| Voice mode | ✅ Local STT (zero API cost), streaming TTS | ❌ |
| Skill format | agentskills.io (open standard) | Proprietary |
| Privacy/security | Zero telemetry, container sandbox | Device pairing, Gateway auth |
| RL training | ✅ Atropos RL training support | ❌ |
| Batch processing | ✅ Batch trajectory generation | ❌ |
| SOUL.md | ✅ Native persona system | Similar but not Slot #1 |
Complementary usage (advanced users)
“Use OpenClaw as the conductor: routing tasks, handling auth, integrating with infra. Use Hermes as the lead specialist: when a task really benefits from long-term memory and learned skills, route it to Hermes via MCP or a custom tool.”
9. Use-Case Decision Tree
需要 AI Agent？
├─ 需要多角色团队隔离 + 企业治理 + 5+ 商业渠道
│ └─ → OpenClaw（Gateway 架构天生适合）
├─ 重复性服务业务（agency/咨询/SaaS ops）
│ └─ → Hermes（自动沉淀工作流为技能，越用越强）
├─ 需要 Signal 或 Feishu 渠道
│ └─ → Hermes（OpenClaw 不支持）
├─ 需要本地零成本 STT（语音交互）
│ └─ → Hermes（faster-whisper 本地运行，无需 API key）
├─ 需要 MCP 外部工具生态（GitHub/数据库/内部 API）
│ └─ → Hermes（原生 MCP 支持，动态发现）
├─ 已有成熟 OpenClaw 团队基础设施
│ └─ → OpenClaw（生态成熟，治理完善）
└─ 需要 RL 训练 / 轨迹导出 / 研究工作流
└─ → Hermes（Atropos RL + Batch Runner 内置）