I have always liked the idea of having a personal knowledge system where notes, ideas, research, projects, and useful information do not simply disappear into folders and browser tabs.
The problem is that taking notes is easy.
Actually using those notes later is the hard part.
That led me to experiment with a different approach: combining Obsidian, as my personal knowledge base, with Hermes Agent, as an AI agent that can work with that knowledge.
The goal is not to build another chatbot.
The goal is to build a Second Brain that I can interact with through an AI agent.
The Problem
Traditional note-taking usually looks something like this:
Idea
↓
Create Note
↓
Forget About It
↓
Search Manually Later
↓
Maybe Find It
Even when the information is stored safely, there is still a gap between having knowledge and being able to use it effectively.
I wanted to explore a workflow where an AI agent could help bridge that gap.
┌─────────────────────┐
│ Me │
│ Ideas / Questions │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ Hermes Agent │
│ │
│ Understands task │
│ Searches knowledge │
│ Creates/updates │
│ notes │
└──────────┬──────────┘
│
▼
┌─────────────────────┐
│ Obsidian │
│ │
│ Markdown knowledge │
│ base │
└─────────────────────┘
This became the foundation of my Second Brain experiment.
Why Obsidian?
I chose Obsidian because the underlying knowledge is stored as Markdown files rather than being locked inside a proprietary database.
That makes the system simple and transparent.
My knowledge lives in a vault, and the vault is ultimately just files.
Obsidian Vault
│
├── Notes
├── Projects
├── Ideas
├── Research
├── References
└── AGENTS.md
This is important for an AI-assisted knowledge system.
The AI should work with my knowledge.
It should not own my knowledge.
Why Hermes Agent?
Obsidian provides the knowledge layer, but I wanted an agent that could actually work with the vault.
That is where Hermes Agent comes in.
Hermes is an AI agent framework from Nous Research that supports tools, skills, context files, persistent memory, and integrations such as Obsidian.
Instead of treating the AI as a simple question-and-answer interface, I can give it a task and allow it to work through the available tools and instructions.
For this project, Hermes becomes the agent layer sitting between me and my knowledge base.
The Architecture
My current setup is intentionally simple.
The main components
| Component | Role |
|---|---|
| Obsidian | Personal knowledge base |
| Hermes Agent | AI agent and orchestration layer |
| Obsidian Skill | Allows Hermes to work with the vault |
| AGENTS.md | Defines how the agent should behave |
| OpenRouter | Model routing layer |
| MiniMax Free | LLM used by the agent |
The interesting part is that none of these components needs to be a giant system.
Each has a specific responsibility.
Hermes + Obsidian
The important part of this project is not simply connecting Hermes to Obsidian.
It is changing the way I interact with my knowledge.
Instead of:
"Where did I save that note?"
the interaction becomes closer to:
"Find the information I previously wrote about this."
Or:
"Create a note from this idea."
Or:
"Update my existing project note with these new details."
The agent becomes an interface for working with the knowledge base.
The Role of AGENTS.md
One of the most useful parts of the setup is AGENTS.md.
An agent needs more than access to files.
It also needs instructions about how those files should be handled.
For example, I can define rules around:
- where different types of notes belong
- how notes should be structured
- how existing notes should be updated
- how new information should be organized
- what the agent should avoid doing
This creates an important separation:
The Markdown files contain the knowledge.
AGENTS.md contains the rules for working with that knowledge.
How a Typical Interaction Works
Suppose I have previously written something about a project.
Later, I ask Hermes a question about it.
The important concept is that the answer can be grounded in my own stored information, rather than depending entirely on what the model already knows.
The agent searches the vault, reads the relevant Markdown files, understands and organizes the context, and then answers using my knowledge.
The Second Brain Is More Than AI Memory
This distinction became important while building the system.
There are actually different kinds of "memory" involved.
Obsidian
Stores the actual knowledge I intentionally keep.
AGENTS.md
Defines how the agent should operate within the knowledge system.
Hermes memory
Provides persistent agent context where applicable.
These are related, but they are not the same thing.
That separation makes the architecture much easier to reason about.
Where the LLM Fits
I am currently using MiniMax Free through OpenRouter as the model layer.
The model provides the reasoning capability.
Hermes handles the agent workflow.
Obsidian holds the knowledge.
That separation is one of the things I found most interesting about the experiment.
Skills: Giving the Agent Procedures
Another useful concept in Hermes is Skills.
A skill is essentially a reusable set of instructions for performing a particular type of task.
This is different from simply giving an AI more information.
Information tells the agent what something is.
A skill can tell it how to do something.
For a Second Brain, that opens interesting possibilities.
Over time, this can turn the Second Brain from a collection of notes into a system with repeatable workflows.
My Current Vault
An important part of this project is that I am not pretending this is already a giant, perfectly organized knowledge graph.
It is still evolving.
My current Obsidian setup is intentionally small, with the agent instructions in place and the structure being developed as I experiment.
That is actually useful.
I would rather prove the workflow with a small knowledge base first than create thousands of notes that have no consistent structure.
The project is therefore less about having a massive vault and more about establishing a good foundation for one.
The Bigger Idea
The system becomes a loop.
I learn something. I capture it. The knowledge becomes part of my Second Brain.
Later, Hermes can help me retrieve or work with that knowledge.
Then the result can become new knowledge.
What I Learned
Building this was less about installing an AI agent and more about understanding how the different layers should work together.
1. AI is not the knowledge base
The model can reason, but it should not be treated as the permanent source of truth.
My notes remain the source of personal knowledge.
2. Instructions matter
Giving an agent access to files is not enough.
Clear instructions such as those in AGENTS.md can make its behavior much more predictable.
3. Start with a small system
A Second Brain does not need thousands of notes on day one.
A small, consistent structure is more useful than a huge collection of disconnected information.
4. The model is only one part of the system
MiniMax provides the language model.
OpenRouter provides the model access layer.
Hermes provides agent behavior.
Obsidian provides persistent personal knowledge.
Each layer solves a different problem.
What I Want to Build Next
This is still an evolving project.
The direction I want to explore is a more capable personal knowledge system where Hermes can help with workflows such as: Capture, Understand, Organize, Connect, Retrieve, Create, Improve, and Capture Again.
The interesting question is no longer:
"Can AI answer my questions?"
It becomes:
"Can AI help me continuously work with the knowledge I have accumulated?"
That is the idea behind my Obsidian Second Brain experiment.
Final Thoughts
This project started from a simple frustration with traditional note-taking.
I did not want another place to store information.
I wanted to experiment with a system where my knowledge could become something I could actively interact with.
Obsidian gives me ownership and structure.
Hermes gives me an agent that can work with that knowledge.
AGENTS.md gives the agent operational rules.
OpenRouter and MiniMax provide the model layer.
Together, they form the foundation of an AI-assisted Second Brain that I can continue to evolve.
It is not a finished system yet.
And that is intentional.
The interesting part is building the system gradually, learning how the pieces interact, and discovering what an AI agent can actually do when it has access to a structured body of personal knowledge.