Hermes is an open-source AI agent that can run entirely on a personal computer. Its chats stay on the local machine, and it stores memory that improves with use. The latest Hermes update adds a guided onboarding process and one-click installation of local models and runtimes. This article covers the installation process on Windows, the onboarding flow, the configuration of a local model, the built-in memory system, key settings, and advanced features for later exploration.

Overview of Hermes

Hermes includes several core capabilities:

• Persistent memory: Hermes remembers projects and user preferences across sessions.
• Automatic skill generation: Hermes creates its own skills and workflows. After it solves a problem, it stores the solution so it can repeat the approach.
• Tool and automation support: Hermes connects to external tools and runs automations.
• Web research: Hermes browses the web to gather information for research tasks.
• Sandboxing: Hermes can run in an isolated environment to protect the rest of the computer.

Hermes is developed by Nous Research. The company works on the open-source movement. It trains models, builds agents, and develops infrastructure to accelerate the adoption of open intelligence globally.

Installation on Windows

The installation process includes two components:

1. The agent, installed directly on the computer.
2. The desktop application, which provides the interface for using the agent.

The installation took about 4 to 5 minutes. After launch, Windows displayed a security prompt asking whether to allow private network access. The author allowed access because he planned to connect this computer to another computer in a later setup. The prompt may also appear on macOS.

Onboarding

On first launch, Hermes presents an onboarding system. Users can complete the process or skip it.

The onboarding runs on a free trial model built into Hermes. This model allows users to complete setup before connecting their own models. Users who prefer to run a local model from the start can bypass the connection to this free model during installation.

The onboarding includes the following elements:

• Connectors: Hermes suggests connections to installed apps and developer tools. Users can review this list later under capabilities and connectors.
• Quick tour: Hermes explains its own structure. The tour covers the following areas:
  • Sessions: Stored chats, shown in the side panel.
  • Chat input: The field for sending messages.
  • Model picker: The control for selecting the model that answers. If the computer can run a model locally, the picker directs the user to settings, providers, and local models.

During the tour, Hermes generated an animated HTML guide that explains how it works. The guide used several skills, including the Hermes agent skill, a design skill, and a p5.js skill.

Agent Loop

The generated guide describes the agent loop. In this loop, a model:

1. Runs terminal commands.
2. Extracts information from the web.
3. Uses skills.
4. Produces a graphical user interface tour.
5. Returns the result as a tool result, which passes back into the conversation.

The model is the component the user controls. Hermes provides the surrounding system that operates on top of the model.

Skills

Skills are recurring workflows that Hermes consolidates. Hermes loads them automatically when they become useful. Four skills loaded during onboarding: the Hermes agent skill, a design skill, the p5.js skill, and a humanizer skill.

Memory Files

Hermes includes two built-in memory files. One stores the work in progress. The other stores how the user prefers to work.

After the onboarding, Hermes suggested running a local model. The suggestion was Qwen 3.6 35B, with 3 billion active parameters. The model requires 22 GB and runs on the author's hardware. The user can decline the paid Nous Research account, which provides access to models such as Claude, ChatGPT, and Gemini.

Configuring a Local Model

Local model setup involves three components: a model, a runtime, and a connection to Hermes.

Model Sources

Hermes supports several provider types. Under Settings > Providers, users can:

• Connect to cloud providers, such as a ChatGPT or Codex subscription.
• Connect to custom endpoints. These endpoints run a model on the same computer or on another device. Examples include Ollama, LM Studio, and Lemonade.
• Use the local model section, which provides model recommendations.

The local model section suggests a model that fits the computer's hardware. Users can select Let me choose to view all compatible models and their fit assessments.

Runtime Installation

A runtime, or inference engine, runs the model and connects it to Hermes. Earlier versions required manual setup of the engine, download of the model, and manual connection to Hermes. In the current version, users can select Install runtime to install llama.cpp directly within Hermes. In the author's test, installation took about 3 seconds.

Model Selection and Memory Fit

The model must fit within the available GPU memory, with additional headroom. The headroom allows the model to retain context as the conversation grows. A model that exceeds available memory cannot run. For example, a 16 GB RAM system cannot run Qwen 3.6 35B, which requires 22 GB.

The author's test computer has 64 GB of unified memory shared between the GPU and RAM.

Model selection involves these concepts:

• Quantization: A method of compressing a model. The letter Q in a model name indicates quantization. A lower-precision version uses less memory. For example, the Q6 version of Gemma 4 12B requires 9 GB, compared with 22 GB for the full version.
• Active parameters: The number of parameters activated during each step. Gemma 4 26B A4B has 26 billion total parameters and activates 4 billion at a time.

Models are downloaded from Hugging Face, a repository of local models. Gemma 4 is one of the most popular models, with 1.6 million downloads.

The author downloaded a smaller Gemma 4 variant that fits within 16 GB of RAM. Users should review each model's hardware requirements before download.

Starting the Model

To start the model:

1. Select Turn on in the local runtime section to start the engine.
2. Select Use next to the installed model. Hermes sets the model as the default.
3. Choose the reasoning setting.
4. Send a message.

On the first message, Hermes loads the model into memory. Loading time depends on model size. After loading, the model keeps running in the background. The author's test loaded 11 GB of the model into RAM.

Exceeding available RAM substantially reduces performance.

Performance

The test model response was produced entirely on the local machine. The model averaged 13 tokens per second. Local models generally run more slowly than frontier cloud models such as Claude Code or ChatGPT. Smaller models often run faster. Quantization level and model size both affect speed and output quality.

Operating Mode and Data Privacy

Local models keep chat content on the computer. Data leaves the machine when the user uses the browser or connects to external tools. Users determine how they structure workflows with local agents.

Use cases for a local agent include:

• Answering questions about private notes.
• Analyzing sensitive company data.

Memory System

Hermes stores information it learns during conversations. In the demonstration, the author asked Hermes to remember that he works on a tutorial YouTube series and prefers short, concise, pragmatic responses. Hermes confirmed that it would save the information to memory.

In a new session, the author asked, "What project am I working on?" Hermes answered that the user works on a tutorial YouTube series that teaches people how AI agents work.

The memory system accumulates information over time. The author states that this allows less powerful models to become more effective for a specific use case, because the stored memories shape each response. The memory functions as a portable profile that can connect to any model.

Key Settings

The author recommends reviewing three settings:

• Memory and context: Confirm that persistent memory and the user profile are enabled. Custom memory providers are available as plugins. For example, users can search for Hindsight under Browse and install it as a memory provider. Honcho is another memory provider option.
• Advanced terminal execution back-end: This setting determines where the agent runs code. Users can run code in Docker. Docker isolates the agent, which protects the rest of the computer if the agent malfunctions or is compromised.
• Messaging gateway: Hermes can connect to Telegram, Discord, or Slack. This connection allows users to communicate with an agent running on their computer from another device.

Additional configuration options include:

• Cloud and local switching: Users can move between local and cloud models.
• Profiles: Users can create profiles with separate configurations. For example, a researcher profile can use settings that differ from the default profile. Profiles are similar to bot mode.
• Bot mode: Multiple bots can communicate with one another.

Advanced Features

The author describes several advanced capabilities covered in other videos on his channel. The following table summarizes them.

The author notes that the memory layer combines Obsidian, chat history, and memory providers into a context stack.

Access to Obsidian can be limited, which restricts the agent's visibility into notes and files. This restriction also uses Docker.

Next Steps

The Hermes setup described in this article uses a local model with a built-in memory system. The author plans future work on:

• Tiers of local models and their fit with specific workflows.
• Hardware requirements for each tier.
• Obsidian integration.
• Advanced memory workflows.
• Bot and profile structures that use these tools with local models.