An agent designed to outlive the session
Most agents reset when the conversation ends. Hermes Agent, from Nous Research, is built on the opposite premise: it is a persistent personal agent with a closed learning loop. It creates skills from experience, refines them during use, searches its own past conversations, and builds a model of who you are that deepens across sessions. The tagline, “the agent that grows with you,” is a design claim, not a slogan, and it is the reason to look past the crowded agent field at this one.
The second defining choice is that it is not tied to your laptop. Hermes runs from a single gateway process that you can reach through Telegram, Discord, Slack, WhatsApp, Signal, or a real terminal UI, while the agent itself works on a cloud VM. You can talk to it from your phone while it runs unattended somewhere cheap.
What the learning loop actually consists of
- Autonomous skill creation after complex tasks, with skills that self-improve during use, compatible with the agentskills.io open standard.
- Agent-curated memory with periodic nudges to persist what matters, rather than dumping everything into a context window.
- Cross-session recall via FTS5 full-text search over past conversations, summarized by the model so old work is retrievable.
- User modeling through Honcho dialectic modeling, so the agent’s picture of you is structured, not just a transcript.
Model-agnostic by default
Hermes does not bind you to one provider. You switch with hermes model across Nous Portal, OpenRouter’s 200+ models, NVIDIA NIM, z.ai/GLM, Kimi/Moonshot, MiniMax, Hugging Face, OpenAI, or your own endpoint, with no code changes. That neutrality matters for an agent you intend to run continuously, because it lets you tune cost against capability without re-platforming.
Install
On Linux, macOS, WSL2, or Termux:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bashOn native Windows in PowerShell:
iex (irm https://hermes-agent.nousresearch.com/install.ps1)The installer brings its own dependencies: uv, Python 3.11, Node.js, ripgrep, ffmpeg, and a portable Git Bash that does not touch any system Git. After it finishes, reload your shell and start:
source ~/.bashrc
hermesFrom there, hermes model picks your provider, hermes tools configures the enabled tools, hermes gateway starts the messaging bridge, and hermes setup runs the full wizard.
Where it runs, and why that is the point
Hermes offers six terminal backends: local, Docker, SSH, Singularity, Modal, and Daytona. The last two give serverless persistence, where the agent’s environment hibernates when idle and wakes on demand, costing almost nothing between sessions. It can also spawn isolated subagents for parallel workstreams and call tools over RPC from Python scripts, collapsing multi-step pipelines into low-context turns. This is infrastructure for an agent you leave running, not a chat you open and close.
The number you should not ignore
Hermes carries an enormous open-issue count, nearly 20,000 as of 2026-06. Read that as two things at once. It signals a very large, very active user base, which a young project with that many reports clearly has. It also signals churn: an agent this ambitious, spanning six backends, many chat platforms, and a learning loop, has a wide surface where things break. If you adopt it, expect to be on a fast-moving release train (calendar-versioned, with v2026.6.5 in June 2026), and pin a version if you need stability for an unattended deployment.
hermes-agent versus other agent projects
| hermes-agent | OpenHands | Letta | |
|---|---|---|---|
| Stars | 189,258 | 76,366 | 23,233 |
| Primary use | always-on personal agent | autonomous software development | stateful memory agents |
| Learning loop | built-in, self-improving skills | task-focused | persistent memory store |
| Reach | CLI plus Telegram, Discord, Slack, and more | IDE and local dev | SDK and server |
Counts are from GitHub as of June 2026. OpenHands targets autonomous coding rather than a general personal assistant. Letta is a framework focused on giving agents persistent memory, which overlaps with one piece of what Hermes does but is a building block rather than a finished assistant. Hermes’s distinct combination is the always-on, multi-platform personal agent with the full learning loop wired in.
Related
Hermes is model-agnostic and can point at a local endpoint served by Ollama. If you juggle its config alongside other coding tools, cc-switch manages those. For what else is climbing, see LLM tooling, the daily digest, and the weekly report.
FAQ
What makes Hermes “self-improving”? It creates skills from completed tasks, improves them during use, and curates its own memory with periodic nudges, so it accumulates capability rather than starting fresh each session.
Do I need a powerful machine? No. It is designed to run from a $5 VPS or serverless infrastructure that costs almost nothing when idle, and you interact with it remotely.
Which models can it use? Any, via hermes model: Nous Portal, OpenRouter, NVIDIA NIM, OpenAI, local endpoints, and more, with no code changes.
Why so many open issues? Nearly 20,000 as of 2026-06 reflects a large active user base and a wide feature surface. Expect frequent releases and pin a version for unattended use.