Hermes AI Agent Explained in Plain English: Architecture, Installation, and Usage

What is Hermes

Hermes is a self‑evolving AI agent developed by Nous Research. The core component is the run\[_agent.py] file, which defines the AIAgent class that handles dialogue, tool calls, context, memory and persistence.

Key Capabilities

Chat Q&A across multiple platforms.

Terminal, file, browser and code manipulation.

Scheduled tasks, automation and parallel sub‑agents.

Long‑term memory that improves with usage.

Support for research training, trajectory generation and reinforcement‑learning environments.

Installation

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

Entry Points

CLI: hermes Message gateway (Telegram, Discord, Slack, WhatsApp, Signal, Email):

hermes gateway

Architecture Overview

Overall Architecture Diagram

Core Modules

run\[_agent.py]

: dialogue loop core. prompt_builder: system prompt construction. model_tools: tool orchestration and dispatch. registry.py: tool registration center. hermes_state: SQLite session storage. context_compressor: context compression. prompt_caching: prompt caching. gateway: multi‑platform message entry.

Layered Architecture

User Layer : CLI and multi‑platform messaging.

Core Layer : Agent runtime.

Capability Layer : Tools, context management, memory.

Bottom Layer : Model API and system environment.

Dependency Flow

Tools register via tools/registry.py, are discovered by model_tools, which generates schemas for the model. The generated tool set is then used by the run\[_agent.py] CLI loop.

Tool Invocation Process

model_tools

discovers all tool modules.

Importing a tool triggers self‑registration to the registry.

Tools are grouped by platform (CLI, Telegram, …).

Schema definitions are generated for model consumption.

The model calls a tool; the tool executes and returns the result.

Context Governance

context_compressor : removes old messages and summarizes intermediate content.

prompt_caching : caches prompts to avoid redundant computation.

Session and Memory

SQLite database with FTS5 full‑text search stores sessions, messages and indexes.

Long‑term memory files reside under memory/[USER.md].

Runtime Flow

Start the selected entry point.

Load .env and .yaml configuration.

Instantiate the AIAgent .

Build the system prompt.

Discover and generate tool rules.

Parse provider/API mode.

Invoke the language model.

If the model requests a tool, execute the tool and feed the result back.

If no tool is needed, return the model’s direct reply.

Persist the session and messages to SQLite.

Platform Support

Telegram, Discord, Slack, WhatsApp, Signal and Email are supported. Windows requires WSL2; Android/Termux follow the official installation steps.