Keep exploring · more guides
Hermes AI plus Obsidian is the most powerful free upgrade you can give an AI agent right now.
Imagine opening your agent and finding it already knows your business.
It knows your clients, your goals, and what you decided last week.
You never explain yourself again.
It files itself away every time it finishes a job.
And there is one rule at the heart of it that almost nobody sets up, which I will show you later.
What you're looking at: the loop every AI user is stuck in. The work ends, the context dies, and tomorrow starts from nothing.
Imagine hiring an assistant.
Every morning they wake up and remember nothing about you.
You explain your business again. Your clients. Your goals.
Every single day.
Hermes can already use tools. It can already finish tasks.
But it does not know your business.
A second brain breaks that cycle for good.
If you can make a folder and type a sentence, you can build this.
There is no code anywhere in it. It is text files and plain English.
Four of them are easy. The fifth is the one almost nobody sets up.
What you're looking at: the whole build, left to right. Everything after this is filling these five boxes in.
That is genuinely it. People make this way too complicated.
What you're looking at: no database, no app lock-in. Six plain text files any agent can open.
It shows your files, links them together, and keeps everything on your own machine.
Because it is only text, every agent you run can read the same brain.
What you're looking at: every agent I run points at the same folder. Swap the model, keep the brain.
Hermes, Claude, Codex, OpenClaw, Grok, Apollo and Jarvis all read the same vault.
What you're watching: my actual vault inside the Agent OS — real paths, real notes, touched hours ago.
Most people only build one. That is why their setup falls apart.
What you're looking at: the messy pile on the left, the tidy shelf on the right. Every session moves things right.
These are the real numbers from my vault, measured today.
What you're looking at: measured from my real vault — 614 notes. The whole thing does not even fit in a context window. One clean page does, easily.
Hermes reads the card first. It only opens the cabinet when it has to.
What you're looking at: the card gets read every time. The cabinet only opens when the exact words matter.
The seven are a system called PARA. The last two are what make it an agent brain.
What you're looking at: the real folder counts from my vault today. Nine folders, nothing else at the root.
What you're watching: those folders on my actual machine, and the note count at the end.
Every note goes in a folder. The root stays clean, so the vault never becomes a junk drawer.
Everything new lands in the Inbox, then gets moved where it belongs. Capture never has to be a decision.
New vault in Obsidian, then hand Hermes the folder path and ask for the nine folders.
~/SecondBrain. Skip it and Hermes builds in its own workspace instead. I made that exact mistake while filming this.What you're watching: the real run. One plain-English sentence, and the whole PARA structure exists.
Who you are, what the business does, your numbers, your voice. Every agent reads this before it answers anything about your work.
What's live right now, and what runs forever. This is the file that makes answers current instead of vague.
What you're actually trying to hit. Without it, your agents optimise for nothing in particular.
It says where everything lives, what to read when you start, and what to write when you finish.
# AI Memory Protocol Source of truth for every AI agent (Claude · Hermes · Codex · Grok · Apollo · Jarvis · OpenClaw) If you only remember one rule: read before you work, write before you leave. ## Why this exists The second brain lives in Obsidian — not in each tool's private memory. Any agent can open this vault and know who I am, what is active right now, decisions already made, and what was built in prior sessions. ## Map — where things live Identity / voice / SOPs 04 Resources/ (start: About Julian.md) Today's intent + notes 01 Daily/YYYY-MM-DD.md Active projects 02 Projects/ Ongoing areas 03 Areas/ Decisions + wins 05 Memories/ Entity wiki Wiki/ (people, tools, concepts) Shared agent log Agentic OS/Memories/YYYY-MM-DD.md Per-agent memory Agentic OS/<Agent>/Memory.md Goals Agentic OS/Goals.md ## Session start — every agent, every time 1. Read 04 Resources/About Julian.md for identity and business context. 2. Read today's notes if they exist: the daily note, the shared agent log, and your own agent folder + Memory.md. 3. If the task touches a known project or area, open its _index.md first. 4. Search the vault for the topic before inventing context. Do NOT dump the whole vault into context — pull only what the task needs. ## Session end — after every build Write durable notes. Prefer append, never wipe history. Always append to: 1. Agentic OS/Memories/YYYY-MM-DD.md — shared log 2. your agent daily log 3. 01 Daily/YYYY-MM-DD.md Then promote what matters: made a decision 05 Memories/Decisions Log.md shipped a win 05 Memories/Wins.md built a project 02 Projects/<Name>/ learned an evergreen 04 Resources/ or Wiki/ changed goals Agentic OS/Goals.md captured something 00 Inbox/ ## Style - Short scannable bullets for agent logs - Wikilinks when the note exists: [[About Julian]] - No secrets, API keys or tokens in the vault - Personal detail stays private ## Agent checklist 1. Session start → read About Julian + today + Memory.md 2. Work → prefer vault facts over guessing 3. Session end → append shared log + promote durable facts 4. Never create root-level scratch notes
What you're looking at: the actual protocol file from my vault, scrolling. One file, and every agent behaves the same way.
New notes go underneath. Old notes never get rewritten. And every agent gets its own folder.
What you're looking at: five agents, five folders, nothing overwritten. On the right, the default — everyone writing over everyone.
That one habit is the whole engine. In a minute I'll show you it running.
Five instructions, inside your protocol file.
What you're looking at: five plain sentences in one file. The fifth is the difference between an agent that knows and an agent that guesses.
Same vault, same rules, two questions. Here is what actually came back.
What you're watching: it reads About Julian first, then today's notes, then the project indexes — and lists all 11 notes it used.
Now the question the vault cannot answer.
What you're watching: no invented client, no invented number. It searched, found nothing, and said so.
Because rule five gives it somewhere true to look, and permission to say no.
That one line is what turns a confident guesser into something you can trust with client work.
Searching everything for every question wastes time and tokens. So the protocol carries a map.
What you're looking at: it walks to the right room instead of searching the whole building. Faster answers, far fewer tokens.
Tell Hermes: every voice note or message marked as a note goes straight into the Inbox.
No title. No tags. No sorting. Sorting is not your job.
What you're looking at: the entire input side of the system. You talk, it lands, you move on.
Nothing is ever left untidy, because no session ends without writing itself down.
What you're looking at: three appends and three promotions, every time a job finishes. That's the whole compile step.
What you're watching: a real agent closing out — appending today's entry to my shared log without touching a line above it.
Scheduled jobs read the vault and tell me what changed, before I sit down.
What you're looking at: five jobs I actually have running. The vault reads itself and reports back.
What you're watching: the real job list on my machine — 6am, 7am, 8am and 6pm, every day.
Those jobs use the quiet hours you're already paying for and never touching.
Drop voice notes in from anywhere. They land in the Inbox like everything else.
How I reach my own vault when I'm away from the machine it lives on.
No laptop open, no session to babysit. The vault keeps growing.
Call notes are gold. Everything below goes straight in the raw folder.
What you're looking at: the input list. Feed it these for thirty days and the answers stop sounding generic.
Style rules, banned words, the structures that work — all in Resources, read before anything gets written.
Feed it the calls and the models you teach. It starts answering like you.
Every account's story, in one place your whole team can ask.
Your voice, written down once, applied to everything after.
Questions that used to mean hours of digging through recordings are now one message.
A real answer, from your own calls, with the source notes named.
The raw material is already organised and linked.
It's in there, linked and searchable.
Run the compile once, then tell Hermes to save what you just did as a skill.
The nightly job runs that skill. You can also run it by hand any time.
--- name: vault-compile description: Turn the Inbox into filed notes and clean wiki pages. Run on a schedule, or by hand any time the Inbox has piled up. --- # Vault Compile ## When to use this Any time new notes have landed in 00 Inbox and the wiki has not caught up yet. The nightly job calls this skill directly. ## The rules that never change 1. 00 Inbox is read-only. Never edit it, never delete from it, never rename anything inside it. 2. Only this skill writes to Wiki/. 3. Every wiki page stays under 150 words. 4. Never invent a fact that is not in a raw note. ## Steps 1. List every file in 00 Inbox modified since the last run. 2. For each one, decide which wiki page it belongs to. Use an existing page if there is one. Only create a new page when the topic genuinely has no home yet. 3. Pull out the key points only — decisions, numbers, objections, names, next steps. Drop the filler. 4. Merge into the page. Keep the existing structure. Never duplicate a point that is already there. 5. Link related pages with [[wiki links]] in both directions. 6. Update Wiki/_index.md if a new page was created. ## Output A five bullet brief: what was processed, which pages changed, what was created, anything ambiguous, anything skipped. ## Edge cases - A note that fits two pages: put it on the closer one, link the other. - A note that is unreadable: leave it, flag it in the brief. - No new notes: say so in one line and stop. Do not touch the wiki. - A page over 150 words: split it, and link the halves. ## Quality gate Before finishing, re-read every page you touched. If a sentence would not help an agent answer a question tomorrow, delete it.
What you're looking at: the skill file, scrolling. Written once, then it is one command forever.
One page per person, company, tool or idea. These are my real counts today.
What you're looking at: the real shape of my wiki. Tools and people dominate, because that's what my questions are about.
What you're watching: the real counts, straight off the folder.
Because every agent appends and none of them wipe, the originals are always underneath.
What you're looking at: nothing in this system destroys anything. That's what makes it safe to experiment with.
Wrong: "This sounds technical. Folders, markdown files, scheduled jobs."
Right: You made folders, typed plain English, and sent voice notes. There is no code anywhere in this build.
Wrong: "I don't have time to maintain another system."
Right: You capture by talking, every session files itself as it closes, and the morning brief tells you what happened.
Wrong: "The next model drops and I have to redo all of this."
Right: The vault doesn't care which model you use. Point the new one at the same folder and it wakes up knowing your business on day one.
Members post their wins every day — agency owners, ecom founders, course creators and solo operators across 38 countries. Real businesses, real numbers, in their own words.
Read the 158-page wins doc →Every conversation gets logged back in, so it gets smarter every day on its own.
What you're looking at: the finished machine. You only ever touch the top-left box.
That last one is the moment you'll get it.
