Hermes Agent¶
Self-improving AI agent by Nous Research with a built-in learning loop — it creates skills from experience, improves them during use, and builds a deepening model of who you are across sessions. MIT licensed, 48k+ GitHub stars.
TL;DR
Hermes Agent is the only major AI agent with autonomous self-improvement: it creates reusable skills from complex tasks, patches them when they're outdated, searches its own past conversations via full-text search, and evolves its prompts/code using genetic optimization (DSPy + GEPA). Supports 15+ messaging platforms from one gateway, 6 terminal backends, and runs anywhere from a $5 VPS to Modal serverless.
Overview¶
| Property | Value |
|---|---|
| Repository | NousResearch/hermes-agent |
| License | MIT |
| Language | Python (≥ 3.11) |
| Stars | 48,700+ (as of April 2026) |
| Creator | Nous Research |
| Version | 0.9.0 |
| Docs | hermes-agent.nousresearch.com/docs |
| Self-Evolution | NousResearch/hermes-agent-self-evolution |
Core Philosophy¶
"The agent that grows with you."
Hermes Agent's differentiator is compounding intelligence — the longer you use it, the better it gets at helping you specifically. Neither Claude Code nor OpenClaw learn autonomously from their own experience in the same way.
Key Features¶
1. Autonomous Skill Creation¶
After complex tasks (5+ tool calls), the agent automatically creates reusable skill documents. These skills are:
- Created without prompting
- Stored as structured markdown
- Searchable and composable
- Automatically improved during subsequent use
2. Skill Self-Improvement¶
Skills are patched in real-time when they are:
- Outdated (API changed, dependency updated)
- Incomplete (missing edge case)
- Wrong (produces incorrect output)
3. FTS5 Session Search¶
Full-text search across all past sessions using SQLite FTS5:
- Search your own conversation history from weeks or months ago
- LLM-powered summarization of search results
- Cross-session recall without manual tagging
4. Honcho Integration — User Modeling¶
Dialectic user modeling that builds a persistent model of who you are across sessions:
- Learns preferences, working patterns, and domain expertise
- Adapts behavior to your evolving needs
- Deepens understanding over time without explicit configuration
5. Multi-Platform Gateway¶
15+ supported platforms from one installation:
CLI, Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Mattermost, Email, SMS, DingTalk, Feishu, WeCom, BlueBubbles, Home Assistant.
6. Terminal Backends (6 Options)¶
| Backend | Use Case |
|---|---|
| Local | Development, personal use |
| Docker | Isolation, reproducibility |
| SSH | Remote server execution |
| Daytona | Serverless persistence (hibernates when idle) |
| Singularity | HPC clusters |
| Modal | Serverless cloud (pay-per-use, near-zero idle cost) |
7. Plugin System¶
Extend Hermes with custom tools, commands, and hooks:
~/.hermes/plugins/
├── my-tool.py # Custom tool
├── my-command.py # Custom command
└── my-hook.py # Event hook
Self-Evolution System¶
The companion repository hermes-agent-self-evolution uses DSPy + GEPA (Genetic-Pareto Prompt Evolution) to automatically evolve:
- Skills (workflow documents)
- Tool descriptions
- System prompts
- Agent code
How It Works¶
graph LR
A[Current Skills/Prompts] --> B[Mutation via LLM]
B --> C[Evaluation]
C --> D{Pareto-Optimal?}
D -->|Yes| E[Keep Variant]
D -->|No| F[Discard]
E --> A
- No GPU training required — operates entirely via API calls
- Cost: $2–10 per optimization run
- Underlying research: ICLR 2026 Oral Paper
- MIT licensed
GEPA Process¶
- Mutate — LLM generates text variants of skills/prompts
- Evaluate — Run variants against test cases
- Select — Keep Pareto-optimal variants (multi-objective: quality, cost, speed)
- Repeat — Evolutionary pressure produces measurably better versions
Architecture¶
graph TB
subgraph Core["Hermes Agent Core"]
AGENT["Agent Loop<br>(Python)"]
MEMORY["Memory System<br>(SQLite + FTS5)"]
SKILLS["Skill Engine<br>(Auto-create + Self-improve)"]
HONCHO["Honcho<br>(User Modeling)"]
end
subgraph Gateway["Multi-Platform Gateway"]
CLI_G[CLI]
TG[Telegram]
DC[Discord]
SL[Slack]
WA[WhatsApp]
MORE[10+ more...]
end
subgraph Backends["Terminal Backends"]
LOCAL[Local]
DOCKER[Docker]
SSH_B[SSH]
DAYTONA[Daytona]
MODAL[Modal]
SING[Singularity]
end
subgraph Evolution["Self-Evolution (Optional)"]
DSPY["DSPy + GEPA"]
EVAL[Evaluation]
SELECT[Selection]
end
Gateway --> AGENT
AGENT --> Backends
AGENT <--> MEMORY
AGENT <--> SKILLS
AGENT <--> HONCHO
SKILLS -.->|Periodic| Evolution
Evolution -.->|Improved| SKILLS
Memory Architecture¶
| Layer | Storage | Scope | Retrieval |
|---|---|---|---|
| Session context | In-memory | Current conversation | Immediate |
| Session history | SQLite + FTS5 | All past sessions | Full-text search |
| User model | Honcho | Cross-session identity | Dialectic modeling |
| Skills | Markdown files | Persistent knowledge | Pattern matching |
Security¶
Security Track Record
As of April 2026, Hermes Agent has zero agent-specific CVEs — in contrast to OpenClaw's 9 CVEs in 4 days (March 2026).
Pricing¶
| Component | Cost |
|---|---|
| Software | Free (MIT) |
| API usage | $15–80/month typical |
| Hosting (VPS) | $5–10/month |
| Self-evolution run | $2–10 per optimization |
| Total typical | $20–90/month |
Daytona and Modal backends can reduce hosting costs to near-zero during idle periods.
Comparison with OpenClaw¶
See Full Comparison.
Key differences:
| Dimension | Hermes Agent | OpenClaw |
|---|---|---|
| Philosophy | Self-improving depth | Universal breadth |
| Skills | Auto-created, self-improving | Human-written, marketplace |
| Memory | FTS5 + user modeling | MEMORY.md + daily notes |
| Channels | 15+ platforms | 24+ platforms |
| Stars | 48k | 360k |
| Security | 0 CVEs | 9+ CVEs |
| Stability | More stable (fewer crash reports) | 3–4 crashes/day reported |
Sources¶
- Hermes Agent GitHub
- Hermes Agent Documentation
- Hermes Agent Self-Evolution
- Awesome Hermes Agent
- Community Docs
- The New Stack — Persistent AI Agents Compared
- Nous Research GitHub Org
Questions¶
Open¶
- Will the GEPA self-evolution approach become standard for all agent frameworks, or is it too complex for mainstream adoption?
- How does Honcho's user modeling compare to Claude Code's CLAUDE.md approach at scale (months of use)?
- What is Nous Research's roadmap for Hermes Agent — is v1.0 imminent?
- Can the self-evolution system be trusted to not degrade skills over time (evolutionary dead ends)?
Answered¶
- Q: Does Hermes Agent require a GPU? — No. Operates entirely via LLM API calls. Self-evolution also uses API calls, not local training.
- Q: Can it work with Claude as the LLM? — Yes. LLM-agnostic; works with Claude, GPT, DeepSeek, Gemini, Ollama, etc.
- Q: How does it compare to OpenClaw? — Hermes maximizes depth of learning (self-improving skills, persistent memory); OpenClaw maximizes breadth of integration (24+ channels, 13k+ marketplace skills).