← Blog · 21 August 2026

Hermes Agent: The Self-Hosted AI Assistant That Learns Your Workflows

Updated 19 August 2026 · by Floris Waller Litjens

Most AI assistants wake up new every morning. They answer today’s question well and have forgotten how they solved last week’s problem. Hermes Agent by Nous Research attacks exactly that: an open-source assistant that lives on your server and is built around one promise. It never forgets how it solved a problem.

Who is behind it

Nous Research has been known for years for its open Hermes language models. In February 2026, the team around co-founder Teknium shipped the Hermes Agent: not another model, but a complete assistant under the MIT license, the most permissive standard in open source. The software is free, the code is open, and by design all data stays on your machine, with no telemetry.

The three-layer memory: the actual difference

Hermes does not just store chat logs. Memory has three levels:

Skill memory. Every solved task is stored as a reusable skill. When a similar task appears, Hermes compares the new outcome with the stored solution and replaces it if the new approach worked better.
Conversational memory. Context across channels and weeks.
User modeling. Your preferences, phrasing and priorities, updated when they change.

The practical effect: the assistant you have in November is measurably better than the one you installed in August, on the same underlying model. For recurring work (weekly reports, data hygiene, research briefings), that compounding is the whole point.

Feature overview

Channels: Telegram, Discord, Slack, WhatsApp, Signal, email and CLI from one gateway process; the list keeps growing.
Install: one terminal command (macOS, Linux, Windows/WSL2) or native desktop apps for all three platforms.
Execution: sandboxed across multiple backends: local process, Docker, SSH, up to serverless infrastructure. Same software from laptop to company server.
Models: your choice; 300+ models via the optional Nous Portal, local models included.
Plain-language scheduling: “every Friday at 4 pm, build the weekly report”, no cron syntax.

What happened this week

Hermes is evolving fast, and two announcements from the past days show it. On 17 August, six days after Grok Bot launched, Nous shipped Bot Mode for Hermes Desktop: persistent, always-on agent instances, a direct answer to the commercial competitor. And since 19 August, an automatic scanner (built on NVIDIA’s SkillEvaluator) checks every skill install before you confirm it: for leaked secrets, PII exposure, injected code (including Unicode smuggling) and license problems. Notably, Nous pointed the scanner at its own bundled skills first and fixed eleven of them.

That second item deserves attention. Skill marketplaces are the ugliest weakness of self-hosted agents: third-party code with broad permissions. A vendor visibly taking that seriously is a genuine reason to choose it.

Strengths and limits

Hermes is the youngest of the three big assistant projects. Its ecosystem is smaller than OpenClaw’s: fewer ready-made skills, fewer tutorials, fewer community answers to niche questions. If you want maximum channel breadth or a giant marketplace, OpenClaw serves you better; if you want zero technical involvement, the managed Grok Bot does. Hermes’ positioning is clear: one assistant that learns your workflows and keeps what it learned, rather than a construction kit for many.

For European businesses, the combination of self-hosting, MIT license and no telemetry is also the strongest GDPR story in this market. The full three-way comparison is here.

Next step

The interesting question is not whether Hermes is good. It is which of your workflows benefit from a learning assistant, and which do not repay the setup. An honest answer to that, including the cases where automation is not worth it, is what our AI automation consulting delivers. For private users, we also set Hermes up end to end at a fixed price.

This article was produced with my own AI assistant: it does the research, I check, cut and take responsibility.