If your Hermes Agent starts every session blank and gives you generic answers, the problem isn’t the model — it’s that it has no memory. Here’s the best memory setup for Hermes Agent, why it’s free, and how I get my agents to maintain it for me.
Short answer
- Best memory for Hermes Agent: a plain-markdown Obsidian vault your agents read before working and write to after finishing.
- It’s free, open source, local, and every agent you plug in can read it — not just Hermes.
- Visualise it as a Memory Galaxy so you can actually see what you know.
- The unlock is the feedback loop: your agents write the notes, so it compounds without you.
What “memory” actually means for Hermes
There are two different things people call memory, and mixing them up is why most setups disappoint.
- Session context — what the agent can see in the current conversation. It’s big, it’s fast, and it disappears the moment the session ends.
- Persistent memory — something written to disk that the agent reads before it starts working and writes to after it finishes. This is what survives.
When people say Hermes forgot everything, they almost always mean they only ever had the first one. Every session starts from zero, you re-explain who you are, what your business does and what you decided last week, and the output comes back generic. That’s not a model problem. It’s a missing layer.
The best memory for Hermes Agent: an Obsidian vault
The best memory for Hermes Agent is a plain-markdown Obsidian vault that your agents read before working and write to after finishing. It’s free, it’s open source, it’s just text files on your own machine, and every agent you ever plug in can read it — not just Hermes.
That last point is the one people miss. If your memory lives inside one tool, it dies with that tool. Mine sits in a folder. Hermes reads it. Claude reads it. Antigravity reads it. When a new agent launches next month, it reads it too, on day one, with full context.
- Free and open source — no subscription, no vendor lock-in, no per-seat pricing on your own notes.
- Plain text — markdown is the one format every model on earth reads and writes fluently.
- Local — your client details and business context stay on your machine.
- Agent-writable — agents don’t just read it, they maintain it. I don’t write markdown. My agents do.
- Linkable — wiki-links between notes give the agent a map, not just a pile.
How the options actually compare
Here’s the honest comparison of what you can use as a memory layer for Hermes.
| Option | What it does | Verdict |
|---|---|---|
| Session context only | Remembers within one chat | Free, zero setup — and gone the second you close it |
| A SOUL.md file | One config file describing who the agent is | Great for personality and rules, too small for a business |
| A flat folder of notes | Text files the agent reads | Works, but no links and no structure — it degrades as it grows |
| An Obsidian vault | Linked markdown notes, agent-maintained | Best overall — free, portable, readable by every agent |
| A vector database | Embeddings + semantic search | Powerful at scale, but heavy setup and you can’t read it yourself |
A vector DB isn’t wrong — it’s just the wrong first step. You want a memory you can open, read and correct with your own eyes. Markdown gives you that. Start there, and add retrieval later if you genuinely outgrow it.
Make it visible: the Memory Galaxy
The upgrade that made this click for me was making the vault visible. Your second brain has always been invisible — folders of notes you never open. So I turned mine into a galaxy I can fly through.
- Stars — every note becomes a point of light, so your knowledge is visible at a glance.
- Constellations — every link draws a line, and the clusters reveal your real focus areas.
- Glow — recently updated notes burn brightest, so your now surfaces itself.
- Fly-through — orbit it, zoom into a cluster, click a star and you’re in the note. You recall by navigating, not by searching.
- Improvements — every note you add makes it richer, and so does every note your agents add.
The galaxy doesn’t care whether your folders are tidy. It shows the real shape of your thinking, and the messy clusters are usually exactly where the good ideas are hiding. You don’t need to organise first. You just need to see it.
The feedback loop is the real unlock
A memory that only you maintain will rot. A memory your agents maintain compounds.
- Your agents read the vault before they start a task, so they open with your context instead of a blank slate.
- They work — keyword research, a client draft, a build.
- They write what happened back into the vault as a new note, linked to the related ones.
- The next session — in Hermes, in Claude, anywhere — starts one step further along.
I layer wearable capture on top of that: Omi takes notes on me during the day, and those get exported into Obsidian. So the vault fills up from conversations I never sat down to write up.
The result is that when I ask Hermes “what did I work on last month?” or “give me keyword ideas based on my memory”, it answers with real dates and real projects, because it’s reading a system that’s been quietly updating itself.
What to actually put in it
Start with the notes that kill the most repeated explaining. Mine cover:
- Who I am, what the businesses are, and what I’m trying to hit this quarter.
- A note per client or project — scope, history, preferences, current status.
- My writing voice, so drafts come back sounding like me instead of like a press release.
- A decision log — what I chose, when, and why, so nothing gets re-litigated.
- Daily logs, written by the agents, of what actually got done.
You do not need a hundred notes to feel the difference. Ten good ones will change the first reply you get tomorrow.
Where memory sits in the wider system
Memory is one of three layers. On its own it’s useful; wired into the other two it’s a different machine altogether.
- The brain — Hermes plus whichever models it thinks with, including free and local ones.
- Mission control — one dashboard with a button for every job instead of a terminal.
- Memory — the vault, shared by every agent in the stack.
If you want the step-by-step setup for the vault itself, I’ve written that up separately in the Obsidian memory guide. For how the whole stack fits together, see Agentic OS for Hermes and how to build an agentic OS.
Want my exact setup? The Agent OS with the Obsidian memory system and the Memory Galaxy view is inside the AI Profit Boardroom — video tutorial, step-by-step guide, the install file, daily updates and weekly coaching calls where we set it up together. If you’d rather start free, grab the free AI course and community or my 200+ AI SEO prompts.
The honest take
Memory tools sound boring. Markdown files in folders is not an exciting sentence. But this is the layer most people skip, and it’s the one that separates a smart chat window from something that actually works for you.
It’s also the cheapest upgrade in the entire stack. Obsidian is free. Markdown is free. Hermes can be run free. The only thing it costs you is the decision to point your agents at a folder and let them start filling it in.
FAQ
What is the best memory for Hermes Agent?
A plain-markdown Obsidian vault that your agents read before working and write to after finishing. It’s free, open source, local, and readable by every agent you plug in — not just Hermes.
Is Obsidian free?
Yes. Obsidian is free for personal use and your notes are plain markdown files on your own machine, so there’s no lock-in even if you stop using it.
Does Hermes have memory built in?
Hermes keeps context within a session, and that works well. What it doesn’t do out of the box is give you a shared, human-readable business memory that every one of your agents reads. That’s what the vault adds.
Do I need to organise my notes first?
No. The system reads the vault as it is. Messy clusters still show the real shape of your thinking, and the agents tidy and link as they go.
Do I have to write the notes myself?
No — that’s the point. I don’t write markdown. My agents create and update the notes after every session, so the vault grows whether or not I sit down to write.
Can Claude and other agents use the same memory?
Yes. Because it’s just a folder of text files, anything that can read a file can read your memory — Hermes, Claude, Antigravity, or whatever launches next month.
How long before it makes a difference?
If you use agents daily, you’ll notice it within a day or two. Give it a week and the replies stop being generic, because the agent is answering from your context instead of guessing.
Do I need a vector database?
Not to start. Markdown plus links covers most people for a long time, and you can actually read and correct it. Add retrieval later if you genuinely outgrow it.
The bottom line
The best memory for Hermes Agent isn’t a product you buy — it’s a folder of markdown your agents own. Put it in Obsidian, let your agents read and write it, visualise it so you can actually see what you know, and every agent you plug in from that day forward starts with your full context instead of a blank page.
About Julian Goldie
I run Goldie Agency, a 7-figure SEO agency, and teach this stuff daily on a 394K+ subscriber YouTube channel. I’ve delivered 240+ client projects on Upwork at a 100% job-success score over 10+ years of ranking sites through every major Google update. If you want the systems I actually run, they’re inside the AI Profit Boardroom, and my link building book is free here.