Hermes Agent

Separate durable memory, session history, and procedural skills so experience can be reused

Hermes stores important environment and workflow knowledge in MEMORY.md and user preferences in USER.md, then injects those bounded stores into the system prompt at the start of a new session. Older conversations remain searchable through SQLite and FTS5 session search, so the agent can retrieve specific past discussions without carrying the entire history in every prompt. Repeated procedures can become agentskills.io-compatible Skills, and /learn can turn local code, documentation, web pages, or a workflow just completed in conversation into a reusable skill. A background self-improvement review can also update memory or skills, making learned state a first-class part of the agent rather than an external notebook.

Sources: [1][2][3]

Decouple where you talk to the agent from where its work actually executes

The messaging gateway exposes the same agent through services such as Telegram, Discord, Slack, WhatsApp, and Signal, so a workflow can continue away from the original terminal. Terminal execution can use local, Docker, SSH, Singularity, Modal, Daytona, Vercel Sandbox, and other backends, letting operators choose between direct access to a workstation and an isolated execution environment. Built-in cron scheduling, isolated subagents for parallel work, and MCP integration extend Hermes beyond one-shot chat into longer-running automation.

Sources: [1][4]

For developers who want an agent to learn their way of working instead of being re-briefed from scratch

Hermes fits coding, research, and operations workflows where the same preferences, environment facts, and procedures repeatedly need to be explained to an AI. It is also relevant when you want to work in a TUI at the desk, send requests from a messaging app while away, and move heavier execution into a cloud sandbox without replacing the agent layer. Support for providers such as OpenRouter, OpenAI, Nous Portal, and custom endpoints makes it possible to keep the learned workflow independent from a single model vendor.

Sources: [1][2][3]

Both execution authority and learned state need governance, especially for unattended gateway deployments

Hermes can execute shell commands and modify files, and the local terminal backend acts directly on the host. Its security documentation recommends isolated backends such as Docker, Modal, Daytona, or Vercel Sandbox for production-style gateway deployments. Dangerous-command approval defaults to a smart mode but can be weakened through configuration, so unattended automation and external messaging should be designed with approval policy, user authorization, and execution isolation together.

Memory is writable by the agent by default. If incorrect assumptions or unwanted profile updates should never silently persist into future sessions, memory.write_approval can require review before saves are applied. Skills can also be created and modified by the agent, so operators should periodically inspect what has become durable learned behavior. Data sent to model providers and external messaging services remains subject to those services' policies. Hermes Agent itself is MIT-licensed.

Sources: [4][2][3][5]