Hermes Agent

NousResearch/hermes-agent — a self-improving AI agent from Nous Research. 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 the user across sessions. Hosts on a $5 VPS, a GPU cluster, or serverless infrastructure (idle = pennies).

🔗 Links

Description

• Repo: https://github.com/NousResearch/hermes-agent
• Docs: https://hermes-agent.nousresearch.com/docs/
• License: open source

Download or use

# Linux / macOS / WSL2 / Termux
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
source ~/.bashrc
hermes              # interactive CLI
hermes setup        # full wizard
hermes gateway      # messaging gateway

Native Windows = no; WSL2 required.

🗒️ Description

🧩 What Hermes actually does

• Real terminal interface — full TUI: multiline edit, slash autocomplete, history, interrupt-and-redirect, streaming tool output.
• Lives where you do — Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI — from a single gateway process. Voice memo transcription, cross-platform conversation continuity.
• Closed learning loop — agent-curated memory + periodic nudges; autonomous skill creation after a complex task; skills self-improve during use; FTS5 session search with LLM summarization for cross-session recall; Honcho dialectic user modeling. Compatible with the agentskills.io open standard (i.e. the Agent Skills / Vercel Skills family).
• Scheduled automations — built-in cron with delivery to any platform. Daily reports, nightly backups, weekly audits — in natural language, unattended.
• Delegates and parallelizes — spawns isolated subagents for parallel workstreams; Python scripts call tools via RPC, collapsing multi-step pipelines into zero-context-cost turns.
• Runs anywhere — 6 terminal backends: local, Docker, SSH, Daytona, Singularity, Modal. Daytona / Modal = serverless persistence — env hibernates when idle, wakes on demand, ~zero cost between sessions.
• Research-ready — batch trajectory generation, Atropos RL environments, trajectory compression for training next-gen tool-calling models.

🧩 Model providers (bring your own)

Nous Portal, OpenRouter (200+ models), NVIDIA NIM (Nemotron), Xiaomi MiMo, z.ai/GLM, Kimi/Moonshot, MiniMax, Hugging Face, OpenAI, your own endpoint. Switch via hermes model — no code changes.

🧩 Migration from OpenClaw

hermes claw migrate (interactive or --dry-run/--preset user-data/--overwrite) imports SOUL.md, MEMORY/USER, user skills, command allowlist, messaging configs, API keys (Telegram/OpenRouter/OpenAI/Anthropic/ElevenLabs), TTS assets, AGENTS.md.

✍️ Reasoning for

The most interesting bits for me:

1. Messaging-first workflow — an agent on Telegram that has context on my repos and answers while I’m walking. Claude Code doesn’t give you that out of the box.
2. Closed learning loop + Honcho user modeling — this is exactly what’s missing from “stateless” agents like plain Claude Code. Every session starts from zero there; here the agent adapts.
3. Modal/Daytona serverless persistence — idle cost ≈ 0. Perfect for fire-and-forget cron jobs like a daily LinkedIn brief for LinkedIn Strategy or a nightly digest from Brain.

Risks:

• No native Windows — I’ll run it on WSL2 or a VPS.
• Self-improving skills sound cool, but DELEGATE-52 is a reminder: after ~20 delegated edits LLMs corrupt 25% of a document. You need checkpoints and --dry-run as default.

Alternatives considered

• Claude Code — stronger at native coding, weaker at cross-platform messaging and memory loop
• Agent Zero — Linux/GUI sandbox vs Hermes TUI/messaging; a different interaction philosophy
• Paperclip — orchestrator for a company-of-agents, Hermes is a single self-improving agent
• Ruflo — agent orchestration platform (98 agents, swarm + federation), Hermes is single-agent (Ruflo is a multi-agent platform with MCP)
• Everything Claude Code — cross-harness perf system; in rc.1 it directly integrates Hermes as the operator workflow layer
• OpenClaw — direct precursor; Hermes has built-in migration (hermes claw migrate)

🔗 Resources

• Nous Research: https://nousresearch.com/
• Honcho (dialectic user modeling): https://github.com/plastic-labs/honcho
• agentskills.io — open SKILL.md standard
• Agent Skills / Vercel Skills — compatible skills ecosystem
• NemoClaw — my self-hosted inference, compatible with Hermes via NVIDIA NIM
• DELEGATE-52 — why self-improving skills require an eval pipeline

---

Template: tool