In one line
Hermes Agent is an open source agent from Nous Research that keeps useful memory between conversations, saves procedures as skills, and schedules recurring work.
Repository: NousResearch/hermes-agent
Are you explaining the same project and pasting the same report format every time? Explore Hermes Agent if you want a successful task to become useful again in the next conversation. It combines terminal conversations and tools with memory, skills, and scheduling.
We find the separation of information useful. Project conventions can go into memory, task procedures into skills, and detailed history into session search. Compare that with keeping everything in one increasingly long conversation, and the purpose of each feature becomes clearer.
For a first trial, inspect a public repository and prepare a structured note. Separate the conditions that should persist from the procedure specific to that task. If your immediate need is a connected coding workflow, compare Superpowers as well.
What can you do with it?
Hermes supports reading files and running commands from terminal conversations, with model selection through hermes model. Its README describes terminal environments including local execution, Docker, and SSH. Available tools and capabilities depend on the environment you connect.
Start with three features for recurring work. Memory carries short project conventions and preferences into the next session. Skills contain procedures loaded when needed. Scheduled tasks can run fresh agent sessions and deliver results to files or configured conversation channels.
Ask three questions: Is this a fact the next session needs? Is it a procedure for a particular task? Does it have a defined time to run? These distinctions help keep a complete reporting procedure out of compact memory and settle inputs before creating a schedule.
Messaging connections are available too. First get one CLI conversation working from request to result. The official getting started documentation recommends establishing ordinary chat before layering on gateway, skills, and scheduled work. Each connection adds another place to understand inputs and delivery.
Installation and the first run
For Linux, macOS, and WSL2, the README CLI installation section documents:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
This executes an installation script, so review it according to your team's software installation policy. Use the official installation guidance for other environments, including Windows.
Reload your shell configuration after installation. The README includes bash and zsh options. For zsh:
source ~/.zshrc
hermes model
hermes tools
hermes
Choose a provider and model with hermes model, then select tools with hermes tools. Prepare authentication and billing with the connected provider. Start with the tools needed for public material so a failed step is easier to locate.
For the first conversation, choose something easy to compare against files, similar to the repository summary in the getting started documentation:
Explain this public repository's main entry points and directory roles.
Include the file paths supporting the explanation. Do not edit files.
Finish with three files I should read next.
Compare the response with the repository. If tools fail, inspect the model connection and enabled tools, then use hermes doctor for diagnosis. Add features after one ordinary conversation succeeds. The installation and setup video linked in the official getting started documentation provides a visual walkthrough.
Scenario 1: Carry project conventions into the next conversation
If every conversation starts with document locations and test conventions, explore memory. Hermes uses MEMORY.md for environment facts and working conventions, and USER.md for preferences. Their contents load at session start; the agent can add, modify, and remove entries through its memory tool.
Use the memory tool to save these conventions for future work:
documents belong in docs, and changes use the repository's existing tests.
Keep personal information and credentials out of memory.
Distinguish saved entries from information kept only in this conversation.
The memory documentation distinguishes a claim of saving from an actual tool call. Verify the saved result and start a fresh conversation to see whether the conventions carry over. On messaging platforms, use /new at a natural task boundary. Restarting a process can continue the same conversation.
Memory is bounded space for essentials. Keep enduring facts there and use session search for detailed history. Update obsolete test commands or retired conventions. A newly saved memory enters the next session's starting snapshot, rather than rewriting the current session's initial context.
Establish whose memory it is. The official documentation warns against two agent processes sharing one Hermes home. Give separate agents separate profiles and keep shared standards in repository documentation. A personal agent's memory should be distinguishable from rules the whole team approved.
Scenario 2: Turn a working procedure into a reusable skill
Suppose you repeatedly inspect public repositories and produce notes covering entry points, setup, tests, and open questions. You may want to reuse the method. Hermes skills are instructions loaded on demand; the agent can create or improve them using task experience and source material.
The skills documentation describes /learn for turning a procedure from a conversation into a reusable skill. Refine the procedure in an ordinary conversation first, then try:
/learn Turn our public repository exploration procedure into a reusable skill.
Input: repository path. Output: entry points, setup, tests, open questions.
Include commands only when supported by that repository's README.
Do not include publishing the result to an external service.
Read input requirements and stopping conditions in the generated instructions. If setup is missing from a README, does the procedure leave a question? Has a path specific to one project become a universal rule? Try another public repository to see where reuse stops being appropriate.
Installed skills can be invoked by their slash-command names and can also load when relevant. Make sure the name and description express the actual role. Before accumulating skills, compare them with instructions already serving the same purpose.
Agent-managed updates bring maintenance work. For a shared procedure, choose someone to read the changes. Keep examples that worked and examples that failed; both help assess the next revision. A procedure learned from one successful task still needs review before broader use.
Scenario 3: Move a recurring report toward scheduled execution
Once a manually requested report has a useful format, explore scheduling. Hermes supports natural-language schedules and cron expressions, with optional skills attached. One possible application is preparing an internal review note from public projects' official release pages.
Do not create a schedule yet.
Read the official release pages for the public projects I specify.
Prepare an internal review note with changes and original links.
First show the tools and output format needed for one run.
After refining the result, establish frequency, timezone, source pages, and delivery location. The scheduling documentation describes creating, listing, and pausing jobs through chat and CLI. Inspect configured jobs with:
hermes cron list
hermes cron status
Scheduled agent jobs require a running gateway. A job appearing in a list does not establish that it has executed. Unpinned jobs can also follow model configuration changes, so establish the model and tools used for the reporting workflow.
Compare scheduled output with manual runs while someone can review each result. Look for unsupported changes, incorrect links, and repeated items. Make sure a source-reading failure is visible through the delivery path instead of resembling a successful report.
External delivery requires separately configured connections and allowed users. Tools available during an interactive conversation may be unable to proceed unattended when approval is needed. Assign an owner to read success and failure results and to pause or repair the workflow before adding more jobs.
Which teams should explore it?
Useful capability: Individual builders and small teams repeatedly returning to the same project can explore compact memory, on-demand procedures, and eventual scheduling. It is particularly relevant when the handover between separate sessions matters.
Operating burden: Someone must maintain memory and skills, review scheduled output, and track model costs. Connecting everything at once makes stale instructions and delivery failures harder to distinguish. Begin with one memory, one skill, and one reviewable result.
Execution permissions: The security documentation covers command approval, user authorization, file writes, and container execution separately. Command approval and operating system isolation serve different purposes. Restrict folders, service connections, and data transmission in the selected environment, beginning with public material.
License: The repository uses MIT. Review conditions for additional skills and extensions, and assess model and external service costs separately. Consider both the convenience of persistent context and your team's ability to review the resulting state.
Original sources and further reading
- Hermes Agent repository
- README snapshot
- Official getting started documentation and linked video
- Memory, skills, and scheduled tasks
- LICENSE
- Continue with the Superpowers development workflow and exploration and editing in OpenCode.
Wrapping up
The Hypemarc Builder Community will keep sharing the AI news and open source projects we explore each week. We'll continue gathering useful resources to share with fellow builders.
- Open chat: AI Builders
- Adoption advice and inquiries: Contact form
- Get the next post: Blog RSS · AI resource library