Hermes Agent: Nous Research's self-improving, memory-keeping daemon gets a native desktop app in the Surface Release
— views
Hermes Agent is Nous Research's MIT-licensed, self-improving AI agent: it writes its own skills, keeps memory across sessions, and runs 24/7 on hardware you control. The June 6, 2026 Surface Release adds a native desktop app for macOS, Linux, and Windows.
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bashWhat it is
Hermes Agent is an open-source (MIT) autonomous agent from Nous Research that runs as a persistent daemon on hardware you control — a laptop, a home server, a Docker container, or a near-free-when-idle serverless sandbox. The repo’s tagline is five words: “The agent that grows with you.” The README’s longer pitch is the actual differentiator: it calls itself “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.”
The growth curve has been remarkable even by agent-hype standards. The project launched in late February 2026; NVIDIA’s RTX AI Garage post (May 13, 2026) reported it had crossed 140,000 GitHub stars in under three months and called it “the most used agent in the world according to OpenRouter.” As of this writing the repo sits around 189,000 stars. It is written primarily in Python, talks to essentially any model endpoint (Nous Portal, OpenRouter, OpenAI-compatible APIs, local models), and reaches you through 20+ messaging channels — Telegram, Discord, Slack, WhatsApp, Signal, Matrix, email, SMS, and more — plus a TUI, a CLI, a web dashboard, and now a desktop app.
The learning loop is the product
Most agents ship with skills; Hermes writes its own. When it works through a complex task or gets feedback, it saves what it learned as a reusable skill — procedural memory in file form, compatible with the agentskills.io format, so skills are portable, shareable, and community-contributed. Skills then get revised during use, not just created once. Around that sits the memory machinery: agent-curated long-term memory with periodic “nudges” that prompt the agent to persist what matters, FTS5 full-text search over its own past conversations with LLM summarization, and a user model that deepens across sessions. A built-in cron scheduler delivers the output of all this to any connected channel, and MCP integration extends the tool surface beyond what ships in the box.
That combination — daemon persistence, self-written skills, cross-session recall, and scheduling — is what separates it from session-scoped coding agents. A coding agent forgets you when the terminal closes. Hermes is designed to be worse on day one and better on day ninety.
The Surface Release: the daemon gets a face
The newest release, v2026.6.5 “The Surface Release” (June 6, 2026), is the project’s mainstream play. The headline item is a native Electron desktop app for macOS, Linux, and Windows with drag-and-drop file support, in-app self-updates, and concurrent multi-profile sessions. A remote gateway lets the desktop app front a Hermes instance running elsewhere over secure WebSocket with OAuth or username/password auth — so the daemon can live on your server while the UI lives on your laptop. The web dashboard has grown into a full admin panel (MCP catalog, messaging channels, credentials, webhooks, memory, pluggable OIDC login), the model picker is now fuzzy-searchable everywhere, a /undo [N] command retracts conversation turns, and the default skill bundle was deliberately trimmed, with niche skills moved to optional installs. The release also ships a full Simplified Chinese translation across the desktop app.
Install and run
# Linux / macOS / WSL2 / Android (Termux)
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
# Windows (native PowerShell)
iex (irm https://hermes-agent.nousresearch.com/install.ps1)It runs locally, in Docker, over SSH, or on Daytona, Singularity, and Modal. NVIDIA’s post pairs it with local models — Qwen-class open weights on RTX workstations or a 128GB-unified-memory DGX Spark — for a fully self-hosted, always-on setup.
Practitioner note
A long-lived, self-modifying agent is a different operational animal from a stateless coding agent, and I would treat it that way from hour one. The accumulated state — memory database plus self-written skills — is the value, so back it up like production data. More importantly, skills the agent writes for itself are code that will execute later with your credentials: review them the way you would review a teammate’s pull request, especially anything it created after browsing the open web. Start it on one messaging channel and a narrow set of permissions, watch what it chooses to persist for a week, and only then widen its reach.
Under-considered angle
Self-improvement relocates the drift problem. With a stateless agent, behavior changes only when the model or prompt changes — both auditable events. With Hermes, behavior also changes every time the agent edits its own skills or curates its own memory, which is to say: continuously, silently, and per-user. That is the feature, but it means two Hermes instances diverge into genuinely different agents within weeks, and no eval run on a fresh install tells you much about your instance. The second-order effect is where the lock-in lands: after months of accumulated memory and self-tuned skills, switching costs live in the agent’s artifact store, not in the model — which makes the portability of the skill format (and the fact that memory is yours, on your disk) arguably the most strategically important property in the whole design.