Definition
Hermes Agent is Nous Research's open-source agent harness. It combines a language model with persistent memory, tools, reusable skills, scheduling, and a messaging gateway. Users can run it on their own machine or a server and reach the same agent from a terminal or supported messaging services.
Users choose the model provider and the execution backend. A description of a Hermes deployment needs both: the model reasons, and the configured machine runs its commands.
Origin and attribution
Nous Research's release history dates Hermes Agent to February 25, 2026. Nous Research develops and publishes the software in its official repository under the MIT license. The release record attributes the product to Nous Research without naming a sole creator.
Nous Research had already released language models named Hermes. Those model checkpoints and Hermes Agent are separate artifacts. A further name collision is HERMES, the Dev-Primitives software-engineering framework introduced by Haibo Jin, Xinjie Li, Peng Kuang, and Haohan Wang in an October 6, 2026 paper. That framework attaches resident models to repository components.
How it works
The model chooses tools within the harness's action loop. The gateway routes messages, while configured terminal backends determine where commands run. Scheduled jobs can start work and deliver results through connected channels.
Hermes describes its learning loop through changes to stored context and procedures. The agent can write durable facts to MEMORY.md, maintain preferences in USER.md, search past sessions, and create or revise skills through skill_manage. These are readable files and runtime operations. Saving a new skill does not itself retrain the model's weights.
A global SOUL.md supplies personality and tone. Current documentation loads it from the Hermes instance's home directory, independently of project instructions such as AGENTS.md.
Scope and limits
Remembering a fact requires a successful memory-tool write. The memory documentation warns that an agent can claim it saved something without doing so. Session boundaries also affect when updated memory enters the prompt. A long messaging conversation and a fresh terminal session can therefore recall information differently.
Command approval, gateway authorization, file guards, and an isolated execution backend have different jobs. Hermes explicitly says its shell-pattern rules are not a complete interpreter or an operating-system sandbox. File-write guards cover specific file tools, while a terminal running as the same operating-system user can reach files through shell commands. Unattended approval policies and permanent allowlists also affect which actions run.
Operational significance
Record the Hermes release, model provider, execution backend, enabled tools, and unattended approval settings. Verify saved memory and completed actions through their stored results. Review reusable skills when they acquire new authority, and use operating-system permissions and restricted environments for containment.
The MIT license covers the harness code. Model weights, inference services, and connected applications have their own terms and data handling.
Distinguish it from nearby terms
- A personal agent describes the relationship and continuing service to one user. Hermes Agent is one implementation.
- OpenClaw and IronClaw are other runtimes with their own storage, gateway, and security designs.
- An agent skill stores a procedure; agent memory stores facts or preferences. Hermes supports both.
SOUL.mdgives identity instructions. It does not establish independent authorization to send messages or execute commands.- Hermes model checkpoints and the academic HERMES Dev-Primitives framework have different developers, release histories, and operating behavior.
Check your understanding
A Hermes agent says it remembered a server address, but a fresh session cannot recall it. What evidence would establish that the memory write happened, and which session and storage settings would you inspect before blaming the model?