Hermes agent Obsidian memory means giving Hermes, Claude Code, Codex and OpenClaw one shared Obsidian vault as their memory, set up with a single copy-and-paste orchestrator prompt, a pointer in each agent and a nightly task that writes a daily log. Do that and your agents stop working alone.
📺 Watch: Hermes Agent Gets UNLIMITED Memory With Obsidian?!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
This is the how-to. I'll show you the problem it fixes, the setup step by step, what to save and what to skip, the one mistake that wrecks the whole thing, and how to fix it when an agent "forgets".
I run Hermes daily inside my Agent OS with a profile per model, and this is the memory setup I use across all of them. The full masterclass and the orchestrator prompt are inside the AI Profit Boardroom ($69/mo).
The problem: your agents work alone
Hermes remembers. Your other agents don't.
Hermes keeps notes between sessions. But Claude Code doesn't know what Hermes did this morning. Codex has no idea what OpenClaw shipped last night. Each one starts every job cold, and you end up re-explaining the same project four times a day.
That's the real cost. Not the tokens. Your time. You become the memory, copying context from one agent to the next.
And even inside Hermes, built-in memory is small on purpose. The Hermes docs set the agent's MEMORY.md at 2,200 characters, about 800 tokens, with a separate USER.md for your profile. It's built for essentials, not for months of project history.
The fix is one place every agent reads from and writes to. That place is an Obsidian vault.
📺 Watch: Hermes Agent OS + Obsidian Is INSANE!
How to set up Hermes agent Obsidian memory
Here's the order I do it in. Each step is small. Together they give every agent the same memory.
Step 1: Create the vault
Make one Obsidian vault for your agents. Structure it with PARA: Projects, Areas, Resources and Archives, plus a place for daily logs. Predictable folders mean agents file things in the right spot. My Hermes Obsidian overview has the full folder layout.
Step 2: Turn on the obsidian skill in Hermes
Hermes ships a bundled obsidian skill. Per the Hermes docs, it reads, searches, creates and edits notes in your vault. The docs also make a useful point: when an agent needs a location on every run of a recurring task, a skill is often a better home than a memory entry, because it loads only when it's needed and doesn't eat into the 2,200-character budget.
Step 3: Run the orchestrator prompt
This is the shortcut. Instead of configuring each agent by hand, you paste one orchestrator prompt and it sets everything up: the vault structure, the instructions for each agent, and the daily log routine. One prompt, all four agents. I share the exact prompt inside the masterclass.
Step 4: Give each agent a pointer
Every agent gets a pointer, a short instruction that tells it where the shared memory lives and to read it before it starts work. Hermes, Claude Code, Codex and OpenClaw each get their own. That's how Claude Code ends up knowing what Hermes did.
Step 5: Give each agent a nightly task
Every agent also gets a nightly task: write a short summary of the day into the daily log. What it worked on, what got decided, what's still open. In Hermes, that's a cron job. The Hermes docs show the same shape in their cron guide, a daily job that saves notes into Obsidian with the obsidian skill, and their kanban guide calls it the long-running journal pattern: same profile, a shared directory and cron.
Step 6: Add the bonus prompts
Once logs are flowing, add a morning brief so you see what happened overnight, and a health check so you know the system is still working. Then point a content agent at the logs to turn memory into posts. More on that in my Obsidian Agentic OS write-up.
Stuck on any of these steps for your own setup? Book a free AI strategy session and we'll sort it.
📺 Watch: Hermes + Jarvis + Obsidian Memory is INSANE!
What to save and what to skip
Not every fact belongs in the same place. Hermes has a few homes for information, and the vault is only one of them. Put each thing where it fits and every agent stays fast.
| Where it lives | What belongs there |
|---|---|
| Hermes MEMORY.md | The agent's essentials: environment facts, conventions, things it learned. 2,200 characters |
| Hermes USER.md | Your profile: who you are and how you like to work |
| A skill | A location or routine the agent needs on every run of a recurring task |
| Obsidian vault | The long record: daily logs, projects, areas, reference notes |
| .env file | Credentials and settings only. Per the Hermes docs, this is not memory |
For the daily log itself, keep it tight. What got done. What got decided. What's open for tomorrow. If a line wouldn't help another agent pick up the work, leave it out.
The biggest mistake: saving everything
This is the one that kills the setup. People get excited and tell their agents to save everything. Every message, every file, every thought.
Don't.
When you save everything, the vault fills with noise. Agents spend their context reading junk instead of the stuff that matters. The useful notes get buried. And the graph view, which should look like a clean map of your work, turns into a hairball.
Memory is only valuable if it's worth reading. One clean nightly summary per agent beats a thousand raw logs. Curate, don't hoard.
The same logic shows up in how Hermes treats its own memory: the built-in files are small on purpose, and the docs push long, recurring context out into skills. My Hermes memory OS guide covers that side in more depth.
If you want a second pair of eyes on what your agents should log, grab a free AI strategy session.
Troubleshooting when an agent forgets
The Hermes docs have a whole section on this, and it's worth working through in order.
- Check the write actually happened. Saying "I'll remember that" is just text. Memory only saves when the model calls the memory tool. Open the files and look: cat ~/.hermes/memories/MEMORY.md and cat ~/.hermes/memories/USER.md. The docs note that small local models, roughly under 30B parameters, often claim a save without making one. Use a stronger model for setup.
- Check the write wasn't held for approval. With write approval on, writes outside the interactive CLI wait for review. Run /memory pending, then /memory approve all.
- Check you're on the right profile. Memory is per profile. A work profile reads from ~/.hermes/profiles/work/memories/, not the default folder. Run hermes profile list to see what exists.
- Check memory is enabled. If memory is switched off in config, the agent can't save anything, whatever it tells you.
- Start a new session. The memory snapshot is frozen at session start. Run /new after a save and the next session sees it.
One more from the docs that matters for shared memory: never point two agent processes at the same Hermes profile. Both write memory and each loads the other's writes, which compounds into a mess. For agents that need to share, the docs say to use an external memory provider. A shared vault folder that each agent reads and writes through its own instructions keeps things separate.
For the full stack this sits in, read the Agent OS guide. And I check which models handle tool calls reliably on Goldie Bench before I trust them with memory writes.
FAQ
Do I need all four agents for this to work?
No. It works with Hermes alone. Every agent you add, whether Claude Code, Codex or OpenClaw, just needs its own pointer and nightly task.
Can I talk to the shared memory by voice?
Yes. In my setup the voice agent reads the same memory, so you can ask what any agent did. See my Hermes Apollo voice guide.
How often should agents write to the vault?
Once a night is the rhythm I use. One short summary per agent keeps the vault readable and avoids the save-everything trap.
Where do I get the orchestrator prompt?
It's in the Agent OS and Hermes Obsidian Masterclass inside the AI Profit Boardroom, with the 30-day AI agent roadmap, daily tutorials, four coaching calls a week, a prompt library and 3,400+ members. Want it mapped to your business first? Book a free AI strategy session.
