Hermes Agent: Nous Research's Self-Improving Agent, and Its Real Relationship With OpenClaw

🌏 中文版

Hermes Agent is Nous Research's open-source (MIT) agent framework. The official one-liner positions it as the only agent with a built-in learning loop:

It's the only agent with a built-in learning loop — it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions. — hermes-agent README

That sentence is the skeleton of the whole project, and the cleanest way to state its difference. Everything else — many platforms, many providers, many backends — competitors can match. The learning loop is where it places its bet.

This post opens the series. The goal is not to restate the official docs, which change faster than any blog post can track, but to capture the trade-offs and the failure modes.

The shape of it

Entry points
  CLI / TUI  ·  Gateway (Telegram/Discord/Slack/WhatsApp/Signal/Email)
  Desktop app  ·  Web dashboard  ·  ACP (VS Code/Zed/JetBrains)
  API server (OpenAI-compatible)  ·  Python library  ·  Batch runner
        ↓
AIAgent (run_agent.py)
  prompt assembly  ·  provider resolution (3 API modes)  ·  tool dispatch
  context compression and prompt caching
        ↓
State                         Tool backends
SQLite + FTS5 (sessions)      7 terminal · 5 browser · 4 web · MCP
MEMORY.md / USER.md           files, vision, TTS…

One detail up front: the README says "40+ tools" while the developer architecture page on the same docs site says "70+ tools, 28 toolsets." Numbers like these have a short half-life in this project, so this series never leans on them — check the reference docs when you need an exact count.

The codebase is no longer a pure Python project either: GitHub currently reports roughly 76% Python and 20% TypeScript, the latter coming from the desktop app, the TUI frontend, and the web dashboard.

The learning loop is four separable mechanisms

"Self-improving" sounds like one thing. It is actually four mechanisms you can disable independently — which is exactly what you need to know when one of them misbehaves:

That last row is the easiest trap right now. Older write-ups (including the first version of this post) describe Honcho dialectic modeling as built in. The docs now place it at plugins/memory/honcho/ — if you don't install it, you don't have it. The same layer offers OpenViking, Mem0, Hindsight, RetainDB, ByteRover, and Supermemory as alternatives.

Skills follow the agentskills.io open standard, which matters for the "how locked in am I" question: skills travel between frameworks.

Three layers people keep conflating

The most common beginner confusion is treating these as one thing:

1. Where you type (CLI, TUI, Telegram, desktop, IDE)
2. Where commands run (local, docker, ssh, modal, daytona, vercel_sandbox, singularity — seven terminal backends)
3. Where inference happens (Nous Portal, OpenRouter, Anthropic, self-hosted Ollama or vLLM…)

The three are fully orthogonal. You can type in Telegram, run commands in a Modal cloud sandbox, and infer on your own vLLM. Which also means: when something breaks, first ask which layer broke. The official debugging order — hermes doctor → hermes model → hermes setup → hermes sessions list → hermes gateway status — is layer-by-layer elimination.

Post 5 covers layer 2; post 3 covers layer 3.

On OpenClaw: a migration path, not a succession

The first version of this post called Hermes "the official successor to OpenClaw." That was wrong, and this revision fixes it.

What is true: Hermes ships hermes claw migrate, and hermes setup auto-detects ~/.openclaw and offers to import. It carries over SOUL.md, MEMORY.md/USER.md entries, user-created skills, the command allowlist, messaging settings, and allowlisted API keys. But OpenClaw is a separate project by a separate team and is still being developed. The README's own community section links HermesClaw, a bridge whose stated purpose is running "Hermes Agent and OpenClaw on the same WeChat account" — which would be pointless if one had replaced the other.

Migration details, and what does not come across, are in post 10. For OpenClaw itself, this site has a full documentation series.

Where the line sits versus Claude Code and LangGraph

Read that table carefully. Claude Code having no learning loop is a trade-off, not a defect: memory lives in files a human maintains, buying predictability at the cost of automatic accumulation. Hermes's auto-created skills and curated memory come with the reverse cost — your agent rewrites its own behavior, which in any workflow that needs reproducibility is a liability rather than an asset. That is precisely why write_approval exists.

When not to pick it

• You only want to call an LLM from code. This is a system, not an SDK. It can be used as a Python library, but you inherit a lot of machinery you won't use.
• You need a team deployment. The design still centers a single owner; the access model is roughly "who is allowed to DM this bot."
• You need strict reproducibility. Self-rewriting skills and memory make "same input, same output" hard. If you go ahead anyway, turn write approval on first.
• You don't want another service to operate. It stays resident, it schedules things, it acts on its own.

The rest of the series

What this series does not cover

Deliberately omitted, because they are feature lists rather than trade-offs and would go stale within months: the Desktop app, Web Dashboard, and TUI; voice mode and the wake word; the image and video generation model catalogs; authoring event hooks; profiles and git worktree isolation; context compression and the details of SOUL.md / AGENTS.md; Mixture of Agents; the egress proxy and managed scope. All of it is in the official docs, which will always be newer than this page.

What this replaces

This post supersedes "Hermes Agent: Nous Research's Self-Improving AI Agent," published 2026-04-05; the old URL /posts/ai/2026-04-05-hermes-agent-intro-en now 301-redirects here. It was fully revised against the upstream README and docs site and rewritten as the opener of a ten-post series. Four misleading claims were fixed: terminal backends went from six to seven (Vercel Sandbox was added); Honcho user modeling moved from built-in to a memory provider plugin; "93% Python" no longer holds (now ~76% Python / ~20% TypeScript); and "Hermes is the official successor to OpenClaw" was wrong — it offers a migration path while OpenClaw continues independently. Command-by-command listings were handed back to the official docs.

References

• Hermes Agent on GitHub
• Hermes Agent documentation
• Nous Research
• Nous Portal
• agentskills.io — open skill standard
• Honcho — user modeling system
• HermesClaw — community WeChat bridge
• Atropos — Nous RL environments