一句话总结

Hermes Agent 是 Nous Research 开源的自我进化个人 AI 智能体：它不只是执行任务，更会从任务中学习 —— 自动创建技能、跨会话记忆、越用越聪明，是目前最接近"永久在线、持续成长"理想的开源 Agent。

项目基本信息

它和普通 Agent 有什么根本区别？

绝大多数 Agent 框架的工作方式是：

收到任务 → 规划 → 执行 → 返回结果 → 会话结束 → 遗忘一切

Hermes 在执行结束后，多了一个关键步骤：

收到任务 → 规划 → 执行 → 评估结果
              ↓              ↓
         提取可复用模式    更新用户模型
              ↓              ↓
         写入技能库    → 下次更快更准 → 越来越强

核心差异： Hermes 不是在重复使用同一个"大脑"，而是在不断扩充这个大脑的经验库。

四层记忆架构

Hermes 的记忆系统是其最核心的创新，分为四个层次：

┌─────────────────────────────────────────────────────┐
│  Layer 4: 外部/可插拔记忆提供商（Hindsight、Honcho 等）│
│  Layer 3: 会话归档搜索（SQLite FTS5 全文检索）         │
│  Layer 2: 情节/技能记忆（.hermes/skills/ 目录）        │
│  Layer 1: 系统提示记忆（MEMORY.md + USER.md）—— 始终在场│
└─────────────────────────────────────────────────────┘

Layer 1：系统提示记忆（总是在线）

• 存储在 ~/.hermes/memories/ 中的 MEMORY.md 和 USER.md
• 每次会话开始时自动注入到系统提示
• Agent 通过 memory 工具自主管理（增删改）
• 有字符限制：记忆区约 2200 字符，用户区约 1375 字符
• 满了就自动整合，保持聚焦

Layer 2：技能记忆（经验结晶）

• 完成复杂任务后，Agent 自动将成功的方法写成技能文档（Markdown）
• 存储在 ~/.hermes/skills/
• 技能可搜索、可复用、会自我改进
• 下次遇到相似问题直接调用已有技能，无需重新摸索

Layer 3：会话搜索（长期回忆）

• 所有 CLI 和消息平台会话存入 SQLite（~/.hermes/state.db）
• 使用 FTS5 全文检索，配合 Gemini Flash 对结果进行摘要
• Agent 按需调用 session_search 工具，不占用常驻上下文

Layer 4：外部记忆提供商（可扩展）

• 推荐：Hindsight（结构化知识提取 + 实体解析 + 反思综合，本地运行，零成本）
• 可选：Honcho（用户建模）
• 在 LongMemEval 基准测试中，Hindsight 是所有提供商中得分最高的

技能系统（Skills System）

技能 vs. 工具

技能的三级渐进加载

• Level 0：只加载技能名称和描述（~3K tokens，支持 40+ 技能）
• Level 1：按需加载完整 SKILL.md 内容
• Level 2：加载具体参考文件

Agent 只在需要时拉取所需层级，保持上下文精简。

技能的可移植性

• 与 agentskills.io 兼容，社区共享技能库
• 支持平台限制和条件加载（如：没有 Firecrawl API key 时自动 fallback 到 DuckDuckGo 搜索）

核心架构内部

系统提示构建（prompt_builder.py）

每次会话的系统提示由以下内容组装：

SOUL.md（个性）
+ MEMORY.md + USER.md（记忆）
+ 技能列表（按需加载）
+ 上下文文件（AGENTS.md、.hermes.md）
+ 工具使用指导
+ 模型特定指令

工具注册表（tools/registry.py）

• 47 个注册工具，分布在 20 个工具集
• 每个工具文件在导入时自动注册
• 统一处理：schema 收集、分发、可用性检查、错误包装

会话持久化

• SQLite（WAL 模式）：支持并发读，单写者
• 每轮对话后写入：会话内容、工具调用、执行结果
• 原始记录存入 JSONL 文件
• Cron 定义单独存储在磁盘

上下文压缩的记忆安全机制

对话变长需要压缩时，Hermes 先触发记忆刷新：

1. 独立调用模型，只提供 memory 工具
2. 给 Agent 机会在压缩前保存重要信息
3. 压缩后，未被标记的内容彻底消失

平台与部署灵活性

支持的消息平台（一个 Gateway 进程管理全部）

Telegram、Discord、Slack、WhatsApp、Signal、Matrix、Email、Mattermost + CLI

支持的终端后端（6 种）

模型无关性

可接入任意模型，一行命令切换，无需改代码：

hermes model              # 交互式切换模型

支持：Nous Portal、OpenRouter（200+ 模型）、OpenAI、Anthropic、Kimi/Moonshot、MiniMax、自托管端点（Ollama/vLLM/SGLang）等

自进化伴侣项目

hermes-agent-self-evolution

使用 DSPy + GEPA 进行进化式自我改进：

• 优化技能、提示词和代码
• 与 Nous Research 的 RL 框架 Atropos 集成
• 可生成数千条工具调用轨迹，用于微调更小、更便宜的模型

最新版本亮点：v0.8.0（2026.4.8）

"The Intelligence Release"

• 🔔 后台任务自动通知
• 🆓 Nous Portal 免费提供 MiMo v2 Pro 模型
• 🔄 跨平台实时切换模型（无需重启）
• 🔒 MCP OAuth 2.1 支持
• ⏱️ 智能不活跃超时
• ✅ 审批按钮（人类监督钩子）
• 🔌 插件系统扩展（可注册 CLI 子命令、请求范围 API 钩子、会话生命周期事件）
• 📝 结构化日志（~/.hermes/logs/）+ 启动时 YAML 配置验证
• 📦 合并 PR 209 个，修复 Issue 82 个

与竞品对比

Hermes Agent vs. OpenClaw

Hermes Agent vs. AutoGPT / LangChain

• vs. AutoGPT：AutoGPT 聚焦任务分解和网络交互；Hermes 聚焦记忆、学习和技能沉淀。可互补。
• vs. LangChain：LangChain 是构建 AI 应用的工具箱；Hermes 是一个完整的 Agent 框架，内置了 LangChain 需要你自己实现的自主技能创建能力。

Hermes Agent vs. AutoResearch（Karpathy）

对 Agent 领域的重要意义

1. "Agent 操作系统" 的雏形

v0.6.0 引入的多 Profile 支持让开发者可以从单个安装运行多个隔离的 Hermes 实例（各有独立配置、记忆、技能、会话）。这已经不是"个人助手"，而是可复用的 Agent 运行时。

2. 技能即知识产权

技能可以共享（agentskills.io）、可以被社区贡献。这创造了一个 Agent 知识库的生态：好的解决方案一旦被一个 Agent 发现，就可以共享给所有人的 Agent。

3. 闭合学习循环的落地验证

Hermes 证明了自主技能创建在实际产品中是可行的 —— 不只是研究论文中的概念。

4. 指向"不断复利"的 Agent 未来

当前的 Agent 更像"工具"：用完即走，不留痕迹。Hermes 的设计目标是让 Agent 变成真正的"数字同事"：越用越了解你，越用越高效，形成能力复利。

局限与注意事项

⚠️ 技术门槛较高

• 需要自管基础设施
• 持久状态管理有一定复杂性
• 比 SaaS 产品的设置更繁琐

⚠️ 处于快速迭代阶段

• 版本更新频繁（每周级别）
• 可能有 Breaking Changes
• 适合愿意跟进前沿的开发者

⚠️ 单一指标的局限（与 AutoResearch 共同的问题）

自我改进依赖可量化的反馈信号，对模糊、主观的任务效果有限。

⚠️ 记忆边界

Layer 1 记忆有字符限制，长期使用需要关注记忆整合质量。

安装快速上手

# 一键安装
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

# 首次配置向导
hermes setup

# 启动
hermes
hermes chat              # 直接进入对话
hermes model             # 切换模型

# 从 OpenClaw 迁移（自动检测 ~/.openclaw）
hermes setup             # 向导会自动提示迁移

我的思考空间

（留给自己的笔记区域，随时补充）

-

相关链接

Sources

• GitHub - NousResearch/hermes-agent
• Hermes Agent Documentation
• Persistent Memory - Hermes Agent
• Architecture - Hermes Agent
• Inside Hermes Agent: How a Self-Improving AI Agent Actually Works
• Hermes Agent Memory System - 3-Layer Explained
• AI 101: Hermes Agent – OpenClaw's Rival?
• Hermes Agent: A Self-Improving AI Agent That Runs Anywhere - DEV Community
• OpenClaw vs. Hermes Agent - The New Stack
• NousResearch/hermes-agent-self-evolution

Finished reading? Leave a comment

Claudesidian: Claude Code + Obsidian Starter Kit

Turn your Obsidian vault into an AI-powered second brain using Claude Code.

What is this?

This is a pre-configured Obsidian vault structure designed to work seamlessly with Claude Code, enabling you to:

• Use AI as a thinking partner, not just a writing assistant
• Organize knowledge using the PARA method
• Maintain version control with Git
• Access your vault from anywhere (including mobile)

Quick Start

1. Get the Starter Kit

Option A: Clone with Git

# Clone with your preferred folder name (replace 'my-vault' with any name you like)
git clone https://github.com/heyitsnoah/claudesidian.git my-vault
cd my-vault

# Examples:
# git clone https://github.com/heyitsnoah/claudesidian.git obsidian-notes
# git clone https://github.com/heyitsnoah/claudesidian.git knowledge-base
# git clone https://github.com/heyitsnoah/claudesidian.git second-brain

Option B: Download ZIP (no Git required)

1. Click "Code" → "Download ZIP" on GitHub
2. Extract to your desired location
3. Open the folder in Claude Code

2. Run the Setup Wizard

# Start Claude Code in the directory
claude

# Run the interactive setup wizard (in Claude Code)
/init-bootstrap

This will:

• Install dependencies automatically
• Disconnect from the original claudesidian repository
• Intelligently analyze your existing vault structure and patterns
• Import your existing Obsidian vault safely to OLD_VAULT/ (if you have one)
• Research your public work for personalized context (with your permission)
• Ask you about your workflow preferences
• Create a personalized CLAUDE.md configuration
• Set up your folder structure
• Optionally configure Gemini Vision for image/video analysis
• Optionally configure Firecrawl for web research
• Initialize Git for version control

3. Open in Obsidian (Optional but Recommended)

• Download Obsidian
• Open vault from the claudesidian folder
• This gives you a visual interface alongside Claude Code

4. Your First Session

Tell Claude Code:

I'm starting a new project about [topic].
I'm in thinking mode, not writing mode.
Please search my vault for any relevant existing notes,
then help me explore this topic by asking questions.

Or use one of the pre-configured commands (in Claude Code):

/thinking-partner   # For collaborative exploration
/daily-review       # For end-of-day reflection
/research-assistant # For deep dives into topics

Folder Structure

claudesidian/
├── 00_Inbox/           # Temporary capture point for new ideas
├── 01_Projects/        # Active, time-bound initiatives
├── 02_Areas/           # Ongoing responsibilities
├── 03_Resources/       # Reference materials and knowledge base
├── 04_Archive/         # Completed projects and inactive items
├── 05_Attachments/     # Images, PDFs, and other files
├── 06_Metadata/        # Vault configuration and templates
│   ├── Reference/      # Documentation and guides
│   └── Templates/      # Reusable note templates
└── .scripts/           # Helper scripts for automation

Key Concepts

Thinking Mode vs Writing Mode

Thinking Mode (Research & Exploration):

• Claude asks questions to understand your goals
• Searches existing notes for relevant content
• Helps make connections between ideas
• Maintains a log of insights and progress

Writing Mode (Content Creation):

• Generates drafts based on your research
• Helps structure and edit content
• Creates final deliverables

The PARA Method

Projects: Have a deadline and specific outcome

• Example: "Q4 2025 Marketing Strategy"
• Create a folder in 01_Projects/

Areas: Ongoing without an end date

• Example: "Health", "Finances", "Team Management"
• Lives in 02_Areas/

Resources: Topics of ongoing interest

• Example: "AI Research", "Writing Tips"
• Store in 03_Resources/

Archive: Inactive items

• Completed projects with their outputs
• Old notes no longer relevant

Claude Code Commands

Pre-configured AI assistants ready to use:

• thinking-partner - Explore ideas through questions
• inbox-processor - Organize your captures
• research-assistant - Deep dive into topics
• daily-review - End of day reflection
• weekly-synthesis - Find patterns in your week
• create-command - Build new custom commands
• de-ai-ify - Remove AI writing patterns from text
• upgrade - Update to the latest claudesidian version
• init-bootstrap - Re-run the setup wizard
• install-claudesidian-command - Install shell command to launch vault from anywhere

Run with: /[command-name] in Claude Code

Staying Updated with /upgrade

Claudesidian automatically checks for updates when you start Claude Code and will remind you to run /upgrade when new features are available.

The upgrade command intelligently merges new features while preserving your customizations:

# Preview what would be updated (recommended first)
/upgrade check

# Run the interactive upgrade
/upgrade

# Skip confirmations for safe updates (advanced)
/upgrade force

What the upgrade does:

• Creates a timestamped backup before making any changes
• Shows you diffs for each file before updating
• Preserves your personal notes and customizations
• Only updates system files (commands, agents, scripts)
• Never touches your content folders (00_Inbox, 01_Projects, etc.)
• Provides rollback capability if needed

Safety features:

• All your personal content is protected
• Complete backup created in .backup/upgrade-[timestamp]/
• File-by-file review and confirmation
• Progress tracked in .upgrade-checklist.md
• Can be stopped and resumed at any time

Vision & Document Analysis (Optional)

With Google Gemini MCP configured, Claude Code can process your attachments directly without having to describe them. This means:

• Direct image analysis: Claude sees the actual image, not your description
• PDF text extraction: Full document text without copy-pasting
• Bulk processing: Analyze multiple screenshots or documents at once
• Smart organization: Auto-generate filenames based on image content
• Comparison tasks: Compare before/after screenshots, designs, etc.

Why this matters: Instead of describing "a screenshot showing an error message", Claude Code directly sees and reads the error. Perfect for debugging UI issues, analyzing charts, or processing scanned documents.

Getting a Gemini API key:

1. Visit Google AI Studio
2. Sign in with your Google account
3. Click "Get API key" in the left sidebar
4. Create a new API key (it's free!)
5. Set it in your environment: export GEMINI_API_KEY="your-key-here"

See .claude/mcp-servers/README.md for full setup instructions

Web Research (Optional)

With Firecrawl configured, our helper scripts fetch and save full web content directly to your vault. This means:

• Full text capture: Scripts pipe complete article text to files, not summaries
• Context preservation: Claude doesn't need to hold web content in memory
• Batch processing: Save multiple articles at once with firecrawl-batch.sh
• Clean markdown: Web pages converted to readable, searchable markdown
• Permanent archive: Your research stays in your vault forever

Why this matters: Instead of Claude reading a webpage and summarizing it (losing detail), the scripts save the FULL text. Claude can then search and analyze thousands of saved articles without hitting context limits. Perfect for research projects, documentation archives, or building a knowledge base.

Example workflow:

# Save a single article
npm run firecrawl:scrape -- "https://example.com/article" "03_Resources/Articles"

# Batch save multiple URLs
npm run firecrawl:batch -- urls.txt "03_Resources/Research"

Getting a Firecrawl API key:

1. Visit Firecrawl and sign up
2. Get 300 free credits to start (open-source, can self-host)
3. Go to your dashboard to find your API key
4. Copy the key (format: fc-xxxxx...)
5. Set it in your environment: export FIRECRAWL_API_KEY="fc-your-key-here"

Helper Scripts

Run these with pnpm:

• attachments:list - Show unprocessed attachments
• attachments:organized - Count organized files
• attachments:sizes - Find large files
• attachments:orphans - Find unreferenced attachments
• vault:stats - Show vault statistics

Advanced Setup

Quick Launch from Anywhere

Install a shell command to launch your vault from any directory:

# In Claude Code, run:
/install-claudesidian-command

This creates a claudesidian alias that:

• Changes to your vault directory automatically
• Tries to resume your existing session (if one exists)
• Falls back to starting a new session
• Returns to your original directory when done

Usage:

# From anywhere in your terminal:
claudesidian

# It will automatically resume your last session or start a new one

The command is added to your shell config (~/.zshrc, ~/.bashrc, etc.) so it persists across terminal sessions.

Git Integration

Initialize Git for version control:

git init
git add .
git commit -m "Initial vault setup"
git remote add origin your-repo-url
git push -u origin main

Best practices:

• Commit after each work session
• Use descriptive commit messages
• Pull before starting work

Mobile Access

1. Set up a small server (mini PC, cloud VPS, or home server)
2. Install Tailscale for secure VPN access
3. Clone your vault to the server
4. Use Termius or similar SSH client on mobile
5. Run Claude Code remotely

Custom Commands

Create specialized commands by saving instructions in .claude/commands/:

Research Assistant (06_Metadata/Agents/research-assistant.md):

You are a research assistant.

- Search the vault for relevant information
- Synthesize findings from multiple sources
- Identify gaps in knowledge
- Suggest areas for further exploration

Tips & Best Practices

From Experience

1. Start in thinking mode: Resist the urge to generate content immediately
2. Be a token maximalist: More context = better results
3. Save everything: Capture chats, fragments, partial thoughts
4. Trust but verify: Always read AI-generated content
5. Break your flow: AI helps you resume easily

Troubleshooting

Claude Code can't find my notes

• Make sure you're running Claude Code from the vault root directory
• Check file permissions
• Verify markdown files have .md extension

Git conflicts

• Always pull before starting work
• Commit frequently with clear messages
• Use branches for experimental changes

Attachment management

• Run npm run attachments:create-organized to set up folders
• Use helper scripts to find orphaned files
• Keep attachments under 10MB for Git

Philosophy

This setup is based on key principles:

1. AI amplifies thinking, not just writing
2. Local files = full control
3. Structure enables creativity
4. Iteration beats perfection
5. The goal is insight, not just information

Contributing

We welcome contributions from the community! This is a living template that gets better with everyone's input.

How to Contribute

1. Fork the repository on GitHub
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Make your changes
4. Test your changes to ensure everything works
5. Commit your changes (git commit -m 'Add amazing feature')
6. Push to the branch (git push origin feature/amazing-feature)
7. Open a Pull Request with a clear description of what you've done

What We're Looking For

• New commands: Useful Claude Code commands for common workflows
• New agents: Specialized agents for specific tasks
• Documentation improvements: Better explanations, examples, or guides
• Bug fixes: Found something broken? Fix it!
• Workflow templates: Share your productive workflows
• Helper scripts: Automation tools that make vault management easier
• Integration guides: Connect Claudesidian with other tools
• Core updates: Improvements to the upgrade system, setup wizard, or other core features

Guidelines

• Keep commands focused and single-purpose
• Write clear documentation with examples
• Test thoroughly before submitting
• Follow existing code style and structure
• Update the CHANGELOG.md with your changes
• AI-generated content is welcome, but you MUST carefully read and review everything before submitting - never submit code you don't understand

Getting Updates

When new features are contributed and merged, users can easily get them with:

/upgrade

The upgrade command intelligently merges new features while preserving your personal customizations, making it easy to benefit from community contributions without losing your work.

Questions or Ideas?

• Open an issue to discuss major changes before starting work
• Join discussions in existing issues
• Share your use cases - they help us understand needs better

Remember: best practices emerge from use, not theory. Your real-world experience makes this better for everyone!

Resources

• Obsidian Documentation
• PARA Method
• Claude Code Documentation

Inspiration

This starter kit was inspired by the workflows discussed in:

• How to Use Claude Code as a Second Brain - Noah Brier's interview with Dan Shipper
• Built by the team at Alephic - an AI-first strategy and software partner that helps organizations solve complex challenges through custom AI systems

License

MIT - Use this however you want. Make it your own.

Remember: The bicycle feels wobbly at first, then you forget it was ever hard.

Finished reading? Leave a comment

这是我的第一篇手记，博客正式开张啦！🎉

（这是初始化占位内容，你可以在后台编辑或删除它，但请始终保留至少一篇已发布的手记。）

Finished reading? Leave a comment