Hermes + Obsidian: a second brain for your agent

Every agent I’ve tried has goldfish brain. Half the session is spent re-explaining the project before I can even ask the actual question. Cool, very productive

Hermes Agent v0.14 fixes this by reading and writing straight into an Obsidian vault. The vault is now the memory.

That’s the whole pitch. Below is the setup and what it actually does once it’s running.

1. The vault IS the memory

No vector DB. No embeddings. No hidden context layer you can’t inspect. Hermes treats your Obsidian folder as plain .md files. It reads them at the start of a session, writes back at the end, and the next session picks up exactly where the last one left off

You can open any of these files yourself and edit them. There is no black box to crack. It’s just markdown, and the markdown is yours.

This is the part that took me a second to internalize. Most “memory” features in agents are opaque on purpose. Hermes is the opposite. You can literally watch the thing think on disk.

2. Install (3 minutes, no cap)

Step 1 -** Install Hermes**

Step 2 - Create vault folder

Step 3 - **Wire Obsidian to Hermes **

Step 4 - Test it

The folder structure does work for you:

• Projects: one note per active project

• Research: scraped pages, sources, references

• Daily: timestamped run summaries

• Memory-Review: candidate facts the agent wants to commit but hasn’t yet

Memory-Review is the part I didn’t expect to ship for. Every “fact” Hermes wants to permanently remember about you lands there first. You read it, accept or delete, and then it migrates to long-term memory. The agent never silently rewrites what it thinks it knows about you. Massively based design choice

3. Config

Two lines and you’re live:

The second one tells Hermes to maintain a profile of you specifically

What you work on, how you write, what you’ve already decided about your own projects. That profile is a .md file too, so when Hermes inevitably gets something wrong about you, you open it and fix it. Same shape as fixing a typo

**4. **What you actually get

Auto-maintained project files. You yap at Hermes about a project and the README, decisions log, and next-actions list update themselves. You can still rewrite them by hand whenever. They’re just notes

A research agent that puts the evidence somewhere real. When you ask Hermes to research something, the scraped content lands in Research/ with source links, not in a chat window that vanishes the moment you close the tab. You can come back a month later and the receipts are still there

Daily briefings. One command in the morning tells you what changed across your projects since yesterday and what’s waiting on you. Morning standup is now zero humans

Graph view, for free. Hermes drops [[wikilinks]] between related notes while it works. Obsidian renders the graph. Watching your agent’s actual cross-references between projects is wildly different from just trusting it.

5. The actual morning command

And the evening one:

6. “But isn’t this just Claude Projects?”

Claude Projects is read-only context you upload. The agent reads what you put there. You’re the one keeping it current

Hermes reads AND writes. The vault evolves while you work. You stop maintaining context for the model, the model starts maintaining its own context. That’s the entire shift

7. Safety, but the actual kind

Start with a scoped folder, not your whole vault. Run the read/write loop for a few days, eyeball what the agent is writing, then expand the path once you trust it. Trusting an agent with your entire knowledge base on day one is ngmi behavior

Secrets stay in .env, never in markdown. Hermes does not read .env. Put keys there. Never paste them into notes

Plain .md plays well with git. iCloud, Google Drive, Syncthing all sync it without complaint. Configure on your server, see the result on your MacBook a few seconds later. No proprietary format to escape from if you decide this tool isn’t for you

Obsidian’s Local REST API is what makes the live read/write work during agent execution. Skip it and Hermes still works, but updates only show up after you reload the vault. Install the plugin

8. What’s under the hood

Three layers under memory.

A 2,200-character durable memory file that survives restarts. The persistent identity layer.

SQLite + FTS5 for full-text search across every past conversation. So “what did I tell you about the auth flow last month” returns the actual answer, not vibes.

Skills: reusable procedures the agent learns and refines over time. Less prompt-engineering, more “the agent already knows how to do this from last week.”

Obsidian sits on top as the human-readable view of all three.

Beyond memory: 22 messaging platforms, 400+ models, 68 built-in tools, MIT license, 100K+ GitHub stars. Runs on your laptop, a VPS, or managed cloud. I run it on a Linux box and pipe everything back to my Mac.

9. Bonus: chain it with NotebookLM

Once your vault is filling up with structured .md, you can point NotebookLM at the same folder.

Hermes writes. NotebookLM reads.

You get audio podcast walkthroughs of your week’s notes and cross-document Q&A across every project at once. Same source of truth, two completely different output modes. Your vault becomes a dataset any AI tool can consume.

This is the part where you realize the vault is the product, and every tool around it is just a different surface for it.

10. Cost

Hermes itself is MIT and runs on your machine for $0. Daily runtime depends on which model you wire it to. Haiku for routine work plus Sonnet for the heavy lifts lands around $0.50 to $2 a day for a solo dev. Local models take it to zero. There is no “Hermes Pro” subscription quietly funneling you anywhere

Why this matters?

I spent way too long pasting the same context into every prompt. Project summary, recent decisions, current blockers, then the actual question. Four-paragraph prompts just to get a usable answer.

That entire ritual was just memory I was carrying for the model

Move it into a vault and your prompts get short again. You stop being the model’s RAM. You start working alongside it instead of briefing it from scratch every session

The leverage is gigabrain