What they are doing
Hermes Agent is an open-source AI assistant built by Nous Research. It combines conversations with tools that can work on files, search the web, use browsers and run scheduled tasks. You choose its model provider and configure the tools it may use. Its distinguishing idea is continuity: carrying useful instructions and experience from one task into later work. Official repository
That continuity has different parts. Persistent memory holds selected facts, preferences and conventions; reusable skills hold task procedures that load when relevant. Hermes can create and revise those skills, and search previous conversations. This is a documented mechanism for saving and retrieving working knowledge; it does not establish that every revision improves the result. Memory documentation, Skills documentation, Project overview
Latest project and updates
The latest tagged release verified by September 20, 2026 was Hermes Agent v0.21.3, released September 14, 2026. It is an available patch release. Its stated priorities include preventing remote desktop and cloud sessions from being signed out during overlapping refresh requests, and correcting duplicate database connections in long-running processes. Those fixes address continuity and reliability rather than introducing a new business application. September 14 release
The larger August 31 release, v0.21.0, added built-in Bot Mode to the desktop app: named agent profiles can participate in group conversations. It also added memory and continuity for scheduled tasks. These are shipped capabilities, although their usefulness depends on configuration and the work assigned. Official release history
Why it matters to your business
Our interpretation: Hermes is interesting where an owner repeatedly explains the same task, corrects the same omission or reconstructs the same report. Three possible applications build on its documented memory, skills and scheduling capabilities:
- Prepare a weekly operations summary using the same headings, exception rules and source files.
- Draft customer follow-up using a saved procedure that checks job status before choosing the message.
- Repeat supplier research while retaining the comparison criteria and recording where each finding came from.
These are proposed applications, not measured customer results. The business value would come from a checked procedure surviving between tasks, with someone responsible for its accuracy. Documented capabilities
Picture this
Consider a fictional repair business preparing invoice-reminder drafts. Its first procedure overlooks partial payments. The owner adds a rule to compare the invoice with the latest payment record before drafting anything. They test three examples: unpaid, partly paid and fully paid. The revised procedure should produce the correct balance for the first two and suppress the third reminder.
In a Hermes setup, that procedure could become a reusable skill. The owner should review subsequent changes and keep the messages as drafts until checked. This illustrates how to use the skills mechanism; it is not a customer case study or a product trial. Skills documentation
Where they are heading
Nous describes Hermes as an agent that develops continuity through memory and reusable skills. The desktop app, plugins and cloud connections broaden how people can access it. The September 14 notes also say that fuller release coverage will accompany v0.22.0, without giving a release date. Desktop overview, September release notes
Our reading is that Nous is moving toward a persistent assistant accessible across several interfaces. That is an interpretation of available features, not a promised roadmap or evidence of unattended business reliability.
Try the idea
Choose one recurring task and save its procedure as version 1. Record one observed mistake, make a small correction and test both an ordinary case and an exception. You can do this with your existing AI application and your own files before deciding whether Hermes is useful.
When the product itself may fit
Hermes may fit an owner or technical helper who wants persistent AI work and can maintain its configuration. The desktop app is free under the MIT license, with downloads for macOS and Windows; Linux installation uses the terminal. Model access, cloud hosting and connected tools can have separate costs. The managed Tool Gateway generally requires a paid subscription; limited free allowances are account-dependent. Desktop availability, Gateway eligibility and billing
It is a poor fit when nobody can review procedures, tool permissions or failures. Skill edits are automatic by default, but an optional approval setting stages them for review. For an initial evaluation, use sample records, enable review of skill changes and restrict file access. Command approval alone is not a complete isolation boundary. Skill-write controls, Security documentation
Explore the original
Sources checked September 20, 2026. Evidence: documentation reviewed; no hands-on product trial.
Sources checked September 20, 2026. Documentation and publisher announcements reviewed; no hands-on product trial. Business examples and resource matches are our interpretation.