Hermes Agent Obsidian Memory: How to Use Obsidian as the Memory Layer

Hermes Agent Obsidian memory means using an Obsidian vault as the main memory layer for Hermes, instead of relying only on its small built-in memory file. Hermes remembers, but its own memory is capped and private to one profile. Your other agents don’t see it at all. This guide is the practical how-to: why the built-in memory runs out, the pointer pattern that fixes it, the daily log format, what to save and what to skip, and how to troubleshoot when Hermes says it remembered something but didn’t.

Why Hermes Agent needs Obsidian memory

Hermes’s built-in memory is good, but it’s small on purpose. According to the Hermes docs:

• MEMORY.md is capped at about 2,200 characters (the memory_char_limit setting), roughly 800 tokens.
• USER.md holds your profile — who you are and how you like to work.
• The memory snapshot is frozen at session start. New memories show up in the next session, so run /new at natural break points.
• Memory is per profile. A work profile keeps its own memories in its own folder, separate from your default one.

That’s fine for preferences and key facts. It’s not enough for months of projects, client notes and research. And it does nothing for Claude Code, Codex or OpenClaw, which can’t read Hermes’s memory. In my video I put it simply: Hermes remembers, but your other agents don’t.

For a broader comparison of Hermes memory options see best Hermes Agent memory and Hermes Memory OS. This page is only about using Obsidian as the layer.

The pointer pattern

The fix is simple. Don’t try to cram knowledge into MEMORY.md. Put a short pointer there instead: where the vault is, which folder to check first, and the rule to write a daily log. The actual knowledge lives in the vault as plain markdown notes.

Every other agent gets the same pointer in its own instruction file. Now they all know where the shared memory is, and the vault can grow as big as you need without hitting the 2,200-character cap.

Set it up step by step

• 1. Make the vault. Install Obsidian and create a vault folder. Note its full path — the docs say workspace paths must be absolute.
• 2. Add the folders. I use PARA: Projects, Areas, Resources, Archive, plus a Daily folder for logs.
• 3. Turn on the obsidian skill. Hermes ships with a bundled obsidian skill that reads, searches, creates and edits notes.
• 4. Add the pointer. Ask Hermes to save the vault path and the log rule to memory. Then check MEMORY.md to confirm it’s there.
• 5. Give other agents the same pointer. Add it to the instruction file of Claude Code, Codex and OpenClaw.
• 6. Schedule the nightly summary. A cron job reads the day’s logs and writes a clean summary note.

In the video I do steps 2 to 6 with one copy-and-paste prompt given to an orchestrator agent. You can do the same, or follow the steps one by one.

A simple daily log format

The daily log is what makes the whole system work. Keep it short. Here’s a simple format you can copy:

# 2026-10-25

## Hermes
- Did: drafted the weekly newsletter
- Decided: send on Tuesdays
- Next: add the new video link

## Claude Code
- Did: fixed the contact form
- Learned: the form needs a honeypot field

## Open questions
- Which offer goes in the P.S.?

Each agent adds a few lines under its own heading. Did, decided, learned, next. That’s enough for any agent to pick up where another one left off.

Nightly summaries with cron

Logs pile up fast. A nightly summary keeps the vault useful. The Hermes docs show the cron pattern with the obsidian skill — their example is a daily paper digest:

/cron add "0 8 * * *" "<your prompt>" --skill arxiv --skill obsidian --name "Paper digest"

Use the same shape for a nightly job: run it late in the evening, use the obsidian skill, and ask it to read today’s log, move decisions into the right project notes, and write a short summary at the top of the log. The docs call this a long-running journal pattern: one profile, a shared folder and a cron job.

What to save and what to skip

The biggest mistake I see is saving everything. A vault full of chat transcripts is as useless as no memory at all, because the agent can’t find what matters.

One tip from the docs: if something happens again and again, it belongs in a skill, not a memory entry. And your .env file is not memory — never put keys in the vault.

Troubleshooting: “I told it to remember”

If Hermes says it saved something but it’s gone next time, the docs list the usual causes:

• Check the write happened. Open ~/.hermes/memories/MEMORY.md and USER.md and look.
• Small local models fake it. Models under roughly 30B parameters sometimes claim a save without calling the memory tool.
• Approval is on. If write approval is turned on, saves wait in /memory pending. Run /memory approve all.
• Wrong profile. Memory is per profile. Run hermes profile list to check which one you’re in.
• Memory is switched off. If memory_enabled is false, the memory tool is removed.
• Old snapshot. Start a fresh session with /new.

Hermes memory vs Obsidian vs a vector database

For most people running a few agents, Obsidian is the sweet spot. You can open it and read exactly what your agents know, fix it by hand, and it costs nothing. A full walkthrough with screenshots is also on the AI Profit Boardroom blog.

My verdict on Hermes Agent Obsidian memory

Hermes’s built-in memory is meant to be small, and that’s fine. Use it for a pointer and your key facts, and let Obsidian hold everything else. Add a daily log, a nightly summary and a clear rule about what to save, and your agents stop forgetting. If you’ve hit the memory limit or want your other agents to share what Hermes knows, Hermes Agent Obsidian memory is the setup to use.

About Julian Goldie

I’m Julian Goldie — founder of Goldie Agency and the AI Profit Boardroom. I test new AI agents almost every day for 432K+ YouTube subscribers and run Hermes, OpenClaw, OpenCode, Claude and Codex side by side in my own Agent OS. Claim your free strategy session or join the AI Profit Boardroom.

FAQ: hermes agent obsidian memory

How much can Hermes Agent remember on its own?

MEMORY.md is capped at about 2,200 characters, roughly 800 tokens, plus a separate USER.md profile file.

Why use Obsidian for Hermes memory?

It holds as much as you need, any agent can read it, and you can open and fix the notes yourself.

Why didn’t Hermes save my memory?

Check MEMORY.md directly, check pending approvals, confirm the profile, and note that small local models may claim saves without calling the tool.

Can two agents share one Hermes profile?

The docs advise against it. Share memory through an Obsidian vault or an external memory provider instead.

Is Obsidian better than a vector database?

For a few agents, usually yes: it’s free, readable and simple. A vector database suits searching very large piles of text.

Related reading

• Hermes Obsidian
• Best Hermes Agent Memory
• Hermes Memory OS

Last updated October 2026. This page is a living guide to hermes agent obsidian memory — the facts here move fast and it is updated as they do.