Most AI assistants forget you the moment the tab closes. Hermes Agent runs as a daemon on your own server, remembers across sessions, and turns the problems it solves into reusable skills.
TL;DR
- Hermes Agent is Nous Research's open-source (MIT) AI agent, released February 2026 and now north of 175,000 GitHub stars.
- It's self-hosted — a persistent process on your own machine, with all data local, no telemetry, no cloud lock-in.
- It has persistent memory across sessions and a self-improving loop: it writes its own skills from experience.
- One agent reaches you on 16+ surfaces — Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI — sharing one memory.
- It's model-agnostic: Nous Portal, OpenRouter, or local models via Ollama/vLLM.
Not a chatbot, not a copilot
The category most AI tools fall into is "ephemeral helper." A chatbot answers and forgets. An IDE copilot is brilliant inside the editor and absent everywhere else. Both restart cold every session, and both run on someone else's servers.
Hermes Agent picks a different shape. It runs as a persistent daemon on infrastructure you control — Linux, macOS, or WSL2 — and stays alive between conversations. It isn't a window you open; it's a process that's already running, accumulating context about your projects and preferences the way a long-tenured colleague would.
The pitch is simple: an agent that lives on your machine and gets smarter every day, instead of a service you rent that resets on every request.
Memory and a real learning loop
Two features make Hermes feel less like a tool and more like a teammate.
Persistent memory. It retains preferences, project context, and past solutions across sessions. You don't re-explain how your repo is laid out or which conventions you follow — it already knows, because it remembered.
Self-written skills. When Hermes works through a non-trivial problem, it can distil the solution into a reusable skill document — a packaged procedure it can reach for next time. Repeated corrections and durable workflow lessons get compacted into memory entries or procedural skills through a consent-aware learning loop. This is the same idea behind agent skills as packaged procedural knowledge, except Hermes authors them itself, from experience, rather than waiting for you to write them.
The result is an agent whose competence compounds. The hundredth task starts much further ahead than the first.
One agent, every surface
Hermes connects to 16+ messaging platforms through a unified gateway — Telegram, Discord, Slack, WhatsApp, Signal, Email, and a plain CLI among them. The point isn't breadth for its own sake; it's that all of those surfaces share one memory and one agent. A briefing you ask for over Telegram on the train draws on the same context as the terminal session you left open at your desk. No context switch, no re-priming.
Built to actually do things
Beyond conversation, Hermes ships with the machinery to act:
- Scheduled automation — a built-in cron scheduler for daily reports, backups, and briefings, set up in natural language.
- Sub-agents — isolated subagents take on parallel workstreams, terminal access, and Python scripts independently.
- Web capabilities — search, browser automation, vision, image generation, and text-to-speech.
- Flexible sandboxing — multiple execution backends (local, Docker, SSH, Singularity, Modal) with container hardening, so you decide how much to isolate what it runs.
- 40+ built-in skills out of the box, on top of the ones it writes for itself.
The local-first bet
The decision underneath all of this is where the agent and its data live. Hermes keeps both on your hardware. Memory, skills, and history sit on disk you own; there's no cloud dependency and no telemetry phoning home. You stay model-agnostic too — point it at Nous Portal's 300+ models, OpenRouter, or a local model through Ollama or vLLM, and swap providers without rewriting your setup.
Self-hosting an agent is the same wager as self-hosting open-weight models: you trade a little convenience for control, privacy, and no lock-in.
Worth watching
Hermes Agent is an early, fast-moving project — but it's a clear articulation of where personal agents are heading: persistent, local, multi-surface, and self-improving. If your mental model of an AI assistant is still "a chat box that forgets," it's worth installing one that doesn't.