Hermes Agent Introduction
Hermes Agent is a self-evolving AI Agent open-sourced by Nous Research. It can continuously learn, self-improve, and run persistently across platforms.
Unlike traditionalChatbotDifferent, Hermes Agent doesn’t just answer questions—it remembers your preferences, automatically distills workflows, maintains context across sessions, and runs synchronously across multiple messaging platforms.
- GitHub repository:https://github.com/nousresearch/hermes-agent
- Hermes Agent Official Website:https://hermes-agent.nousresearch.com/
What is Hermes Agent?
Hermes Agent is a self-evolving AI Agent framework built by Nous Research, released under the MIT open-source license.
The core positioning of Hermes Agent is notChatbot, but rather an autonomous agent—an intelligent agent that runs continuously in the background, learns proactively, and works across platforms.
The official definition is very straightforward:
The agent that grows with you. 一个会随着使用不断成长的 Agent。It is an autonomous agent that becomes more capable the longer it runs.
The predicament of most current AI tools: the amnesia problem
Anyone who has used AI tools has experienced this — spending an afternoon teaching Claude or ChatGPT the special conventions of your project, then closing the chat window after an extremely productive collaboration.
The next day, you open a new session and everything is back to square one.
this is calledThe amnesia problem of AI tools(The Amnesia Problem) — Capability is there, memory is not.
Hermes's solution: a closed learning loop
Hermes solves the amnesia problem through three core mechanisms, forming a self-reinforcing closed loop:
任务执行
↓
提炼可复用技能（Skill）
↓
更新持久记忆（Memory）
↓
下次执行类似任务时自动加载
↓
持续改进 ……
- 1. Persistent Memory: Three-layer memory architecture — working memory (current session), semantic long-term memory (
memory.md), and episodic memory (SQLite + full-text search). The Agent actively organizes and updates its own memory files, instead of relying on you to manually provide context every time. - 2. Autonomous Skill Creation: When the Agent completes a complex task (usually involving more than 5 tool calls), it automatically distills this workflow into a reusableSkill file(Markdown format). The next time it encounters a similar task, it directly loads this Skill without needing to fumble through the process again.
- 3. Model-Agnostic Design: Hermes is not bound to any specific model. It can integrate with Nous Portal, OpenRouter (200+ models), OpenAI, Anthropic, or locally deployed Ollama models—switch with a single command, no code changes needed.
In short:ChatbotIt answers one question at a time; Hermes Agent is given a goal and figures out how to accomplish it on its own.
Comparison with other AI tools
The table below compares Hermes Agent with current mainstream AI tools to help you understand the differences in their core capabilities.
| Comparison Dimension | ChatGPT / Claude.ai | Claude Code / Cursor | Hermes Agent |
|---|---|---|---|
| Cross-session memory | Limited/None | None | ✅ Three-layer persistent memory |
| Auto-learn skills | ✗ | ✗ | ✅ Automatic Skill distillation |
| How to run | Cloud SaaS | IDE Plugin | Local long-running processes |
| Multi-platform access | Web/App | Inside IDE | Telegram/Slack and 15+ other platforms |
| Data Privacy | Upload to Cloud | Upload to Cloud | ✅ 100% local |
| Scheduled automation | ✗ | ✗ | ✅ Built-in Cron scheduling |
| Model Binding | OpenAI/Anthropic | OpenAI/Anthropic | ✅ Any model |
| Open source | ✗ | Part | ✅ MIT License |
Differences from other open-source Agent frameworks
In addition to the above tools aimed at end users, there are many Agent frameworks in the open-source community, and Hermes is fundamentally different from them as well.
Differences between Hermes and AutoGPT / CrewAI
AutoGPT, CrewAI and other frameworks excel atOrchestrate multi-step tasks, but they are essentiallyStateless—each run starts from scratch, with no memory accumulation and no mechanism for distilling workflows from past experience.
What makes Hermes different isTime Dimension: It not only executes the current task, but simultaneously builds an increasingly powerful self during execution.
The difference between Hermes and OpenClaw
OpenClaw is currently the closest competitor to Hermes; both support persistent memory and cross-platform access.
The core difference lies in:
- OpenClawwithControl Plane FirstAs its design philosophy, it emphasizes ecosystem breadth, with skills manually written and maintained.
- HermeswithSelf-evolving loopAt its core, skills are autonomously created and improved by the Agent, emphasizing depth of learning.
In short: OpenClaw is an assistant you carefully train, while Hermes is an apprentice that grows on its own.
Framework comparison at a glance
| Features | Hermes Agent | AutoGPT | CrewAI | OpenClaw |
|---|---|---|---|---|
| Memory System | three-layer architecture (working + semantic + episodic) | Vector database | No built-in memory | No built-in memory |
| Skill Refinement | Automatic distillation + /learn command | None | None | None |
| Multi-platform access | Telegram/Discord/Slack/WhatsApp/Signal | None | None | Telegram/Discord |
| Deployment Mode | 6 types (CLI/Docker/Serverless, etc.) | Docker | Python Script | Docker |
| MCP Support | Native support | Limited | Limited | Limited |
| Scheduled Task | Built-in Cron + hosted Chronos | None | None | no built-in |
| Plugin System | Middleware + Observer Hooks | Limited | Limited | Limited |
| Open Source License | MIT | MIT | MIT | MIT |
Hermes's three-layer architecture refers to: Agent Core (decision-making and reasoning) → Gateway (message routing and platform adaptation) → Connector (platform encryption and identity boundaries). This layered design allows each layer to be independently extended and replaced.
Hermes' core philosophy
Hermes Agent's design revolves around a core closed loop: memory → skills → task execution → self-improvement.
This closed loop allows the Agent to keep growing during use, understanding you better the more it is used.
Memory
Hermes adopts a three-layer memory architecture, covering different time dimensions from short-term to long-term.
Working memory handles the context of the current session, semantic long-term memory (memory.md) stores key facts and preferences, and episodic logs (SQLite + FTS5) record all historical conversations and operations.
The Agent can search and recall prior memories across sessions, meaning it won't forget things you've told it before.
Skills
Skills are Hermes's most core feature — they are an abstraction of procedural memory, driving reusable workflows in the form of Markdown files.
When the Agent performs more than 5 tool calls, it can automatically distill this process into a Skill.
You can also use/learnThe command generates Skills from documents, API documentation, or historical conversations.
The community shares and installs Skills through Skill Hub (agentskills.io), creating an open skill ecosystem.
Task Execution
Hermes has 70+ built-in tools, covering 28 tool collections including file system, web browsing, code execution, image recognition, audio processing, and more.
Through the MCP (Model Context Protocol), you can connect to any third-party tool server to infinitely expand the Agent's capability boundaries.
Self-Improvement
After every 15 tasks, Hermes triggers the Periodic Nudge mechanism to automatically evaluate and optimize its own memory and skills.
The Honcho user modeling system builds a dialectical user profile, allowing the Agent to construct a deep understanding of you across sessions.
Core features
Core features:- Connect:Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI — one Agent, unified memory, all platforms. A conversation started on Telegram can seamlessly continue in the terminal.
- Memory (Remember):Hermes learns your projects, automatically generates skills, and never forgets the problems it has solved. After every session, important information is written to persistent memory.
- Scheduled (Schedule):Set scheduled tasks in natural language—daily reports, backups, routine reviews, morning briefings—running unattended in the background.
- Delegate:Generate isolated sub-Agents with independent conversation contexts, independent terminals, and Python RPC scripts, enabling parallel pipelines with zero context cost.
- Search & Multimodal (Search):Web search, browser automation, visual understanding, image generation, text-to-speech, and multi-model reasoning—all built in.
Hermes Agent supports freely switching between any large model, includingNous Portal, OpenRouter (200+ models), OpenAI, GLM, Kimi, MiniMaxetc., executehermes modelto switch instantly, no code changes, no vendor lock-in.
| provider | Description | Setting Method |
|---|---|---|
| Nous Portal | Subscription-based, zero configuration | Throughhermes modelLog in with OAuth |
| OpenAI Codex | ChatGPT OAuth, using Codex model | Throughhermes modelAuthenticate via device code |
| Anthropic | Directly use Claude models | Via Claude Code authentication or an Anthropic API key |
| OpenRouter | Multi-provider routing | Enter your API key |
| DeepSeek | Direct DeepSeek API access | SettingsDEEPSEEK_API_KEY |
| Hugging Face | Access 20+ open models through a unified router | SettingsHF_TOKEN |
| Custom endpoints | VLLM, SGLang, Ollama, or any OpenAI-compatible API | Set base URL and API key |
| Features | Capability Description |
|---|---|
| Native terminal interaction | Full TUI interface, supporting multi-line editing, command completion, history recall, streaming output, etc. |
| All-platform access | aGatewayJie入 CLI、Telegram、Discord、Slack、WhatsApp waitmanyend |
| Closed-loop learning system | Autonomous memory management, skill generation and optimization, cross-session recall, user modeling |
| Scheduled automation | Built-in Cron scheduling, supporting 7×24 automatic tasks such as daily reports, backups, and audits. |
| Parallel task processing | Supports parallel execution of sub-Agents, multi-workflow splitting, and RPC tool calls. |
| Multi-environment operation | Supports 6 backends including local, Docker, SSH, Daytona, and Modal |
| Research-grade capability | Supports trajectory generation, reinforcement learning environments, and training data compression. |
Overall Architecture
Overall process:
Architecture Diagram:
| Module | Function | Example |
|---|---|---|
| Access layer (Clients) | Receive user requests | Web, App, API, Feishu |
| Input layer (Input) | Process various input data | Text, images, PDF, Excel |
| Scheduler (Agent Orchestrator) | Understand tasks and break down execution workflows | Analyze requirements → Formulate a plan |
| Capability Layer (Capabilities) | Provide execution capabilities | Conversation, retrieval, code execution |
| Memory layer (Memory) | Save context and historical information | Session memory, long-term memory |
| Knowledge layer (Knowledge) | Provide knowledge retrieval capability | RAG, vector databases |
| Model Layer | Provide reasoning capabilities | GPT, Claude, local models |
| Tool layer (Tools) | Call external tools to complete tasks | Search, databases, APIs |
| External Services Layer | Integrate with business systems | ERP, CRM, cloud services |
| Infrastructure Layer | Support system operation | Permissions, logging, monitoring |
Applicable scenarios at a glance
Hermes Agent has a very wide range of application scenarios. Here are a few typical usage directions.
| Scenarios | Description | Typical Configuration |
|---|---|---|
| Personal productivity assistant | Manage schedules, handle emails, organize information, automate daily tasks | Default Profile + Schedule/Email Skill |
| Programming Assistant | Code review, bug fixing, project scaffolding generation, documentation writing | Coding Profile + code-related Skills |
| Research Agent | Literature search, data collection, information organization, report generation | Research Profile + Search/Analysis Skill |
| Automated operations | Server monitoring, log analysis, automatic alerts, scheduled inspections | Ops Profile + Cron scheduled tasks |
| RL training data generation | Batch trajectory generation, ShareGPT format export, integration with Atropos | batch_runner + training Profile |
| Multi-platform customer service | Unified Telegram/Discord/Slack access, with automatic replies to common questions. | Gateway Multi-platform + FAQ Skill |
other extensionsA single Hermes Agent instance can create multiple personas through the Profile system. Each persona is independently configured with its own memory, skills, and tools, without interfering with one another. For example, one as a coding assistant, one as a research assistant, and one as an operations monitor.