How I Use Obsidian as My Context Vault for AI

TL;DR: An AI only knows what is in front of it, and by default that is nothing about you. I keep my notes in Obsidian, a free app for linked plain-text notes, and let my AI read them so I never have to give it the full spiel again. It has become the foundation of how I work with AI, at work and at home, and you can start with three notes in about 15 minutes.

Meeting the same helpful person for the first time, every time

Imagine you meet someone incredibly helpful. They are sharp, patient and they know a lot about a lot of things. You explain your situation and who you are in detail, they give you great advice, and you walk away feeling like you finally have someone in your corner.

The next day you go back and they have no idea who you are. So you explain again. Who you are, what you do, how you like things done. Great advice again. The day after that, same thing.

That is what working with AI was like in the early days. It is a super helpful person you meet for the first time, every single time. This is less true today than it used to be with AI having built-in memory, but at the end of the day, there are many benefits to owning your own context data. This is how Obsidian can help.

How AI context windows work

Context windows are how much an AI can keep in mind at one time. Large language models (LLM), the ones you run through ChatGPT, Claude, and Gemini, can only work with the text that it has access to. That text is its context window. It includes your messages, its own replies, any files you attach and any instructions you give it. Anything outside the window does not exist as far as the model is concerned. That is still true even with built-in memory, which works by loading saved notes into the window for you.

The window is measured in tokens, which are small chunks of text. As a rough rule, 1,000 tokens is about 750 words. Today’s models can hold anywhere from tens of thousands of tokens to over a million, depending on the model and the plan. That sounds like plenty, but it is finite, and when a long chat fills up, older details can get dropped or blurry.

The model itself also does not learn from your chat. When you open a new conversation, the window starts empty. Some products add a memory feature on top, which saves short summaries of past chats. Those help, but you’re not actively controlling what those short summaries look like.

LLMs are quite powerful these days and continue to get better every day. For the average person’s use case, we don’t really need to care or wait for AI to reach Singularity. The question is not how to make the AI smarter. It is what should you put into the context window to make AI work for you the way you want- mimicking intelligence/sentient thought and knowing what to do with your information, as long as it’s got the necessary context on you and your issues at hand. People call this context engineering, and I think it matters more than clever prompts. A great prompt on an empty window still gets you a generic answer.

Here is an example. Say you ask an LLM, “Can you help me plan a trip to Singapore for my family of four next February?” With an empty window, it can only give you the greatest hits: Gardens by the Bay, Marina Bay Sands, the zoo and a hawker center or two. Nothing wrong with that, but it could be anyone’s itinerary.

Now imagine the AI has already read a few notes about you. Your kids are 6 and 9. Your family would rather do hands-on, outdoorsy things than shop. You love spicy food but your youngest does not. And your anniversary is February 20. Same question, very different answer:

Is this a celebration trip? Your anniversary is February 20, so I’d keep that evening free for an adults-only dinner and ask your hotel to arrange something special, like a note or a cake. With kids who are 6 and 9, I’d put the Night Safari and Bird Paradise early in the week and keep the afternoons light. For food, I’d pick hawker stalls with a mild option for your youngest next to a spicier one for you. Want me to draft a note to your hotel letting them know you are celebrating your anniversary?

That answer did not come from a smarter model. It came from a model that knew who it was talking to because you gave it access to your Obsidian vault, and you did not have to type any of it.

Life before Claude, life after Claude

I often joke that 2026 has been pivotal because I learned there is life before Claude, and life after Claude, and life after Claude has been so fun and exciting for a curious systems thinker like me. I signed up for Claude Pro at the end of March 2026, right before I left for a family trip to Singapore. During that same trip, my work got access to Claude Enterprise too.

Before that, I had been building Gemini Gems and vibe coding task management apps. Useful, but with Claude I wanted something bigger. I wanted a system I could really plug into to level up my personal and professional life, get time back and do more with the time I have.

I very quickly saw that prompt engineering was not it on its own. Who wants to give AI a whole spiel every time they need it to do something? I wanted it to already know who I am, what I do and how I like to get things done.

I had read Building a Second Brain by Tiago Forte back in 2022, before ChatGPT (3.5) even came out, and one idea stuck with me: my mind is made for thinking, analyzing and learning, not for storing data. So I set out to build that second brain, and this time it would be one my AI could read too.

Notion and Confluence were the obvious places to start, but both were more complicated than what I needed. I just wanted notes that link to each other. That is how I found Obsidian.

How Obsidian Works

Obsidian is a note-taking app that is free to download. At its core, Obsidian is essentially just a folder on your desktop. Every note is a plain-text markdown (.md) file sitting in a regular folder on your computer, and that folder is called a vault. There is no special format and no account to create. Add [[double brackets]] around a name inside a note and it links to another note, so your ideas connect the way they do in your head.

Plain text is the part that matters for AI. A tool like Claude can be given access to a folder on your computer, which means it can read your notes directly. It pulls in what is relevant to the task, so the right context lands in the window before it starts helping.

It also works the other way around. AI can write to your vault. It can update a note, add to your daily log, or file a new idea next to related ones. That is what turns the vault into a living repository instead of a pile of files you have to maintain by hand.

Why not just use built-in memory?

Fair question, since ChatGPT, Claude and Gemini now offer memory features, and for casual use they are a great zero-effort option. I still keep a vault for four reasons.

First, I can see it. A vault is a set of notes I wrote or approved, so I know exactly what my AI knows about me. Built-in memory is a summary the product decides to keep, and it can miss the details that matter.

Second, I can structure it. I can keep a decision in one note, a project in another, and link them, so the AI gets the right pieces for the right task instead of one blob of facts about me.

Third, it goes everywhere. Memory lives inside one product. My vault works with whichever tool I point at it, so even though I love Claude right now, I still get to keep my options open and am free to change my mind because I can point the next LLM to my Obsidian Vault and it would be able to catch up without any downtime.

Fourth, it is bigger than a profile. Memory is good at remembering that I prefer short answers. A vault can hold my projects, my people, my past decisions and my to-do list, so the AI can help with the actual work.

The graph view

One of my favorite parts of Obsidian is the graph view, which draws your vault as a map. Every note is a dot and every link between notes is a line, so you can see how your ideas, people and projects connect. Notes with a lot of links grow into big hubs, and the lonely dots show you what you have not yet connected to anything.

It is more than a pretty picture. The links that make the map are the same links your AI follows to find related context. When a project note links to a person, a decision and a goal, the AI can pull in all four without you pointing at each one.

How mine is organized

My personal vault is organized by area of life, with a number in front of each folder name so they stay in order. The 000 folder holds the behind-the-scenes things like templates, scripts, my inbox and the instructions I give my AI, and 900 is an archive for things I am done with. Here is the top level:

000 Meta
100 Daily Notes
200 Personal
300 Family
400 People
500 Work
600 Projects
700 Finance
800 Travel
900 Archive
CLAUDE.md
HOME.md
TASKS.md

Three files sit at the top next to the folders. HOME.md is my home dashboard, TASKS.md is my running task list, and CLAUDE.md is the one I care about most. Nothing fancy.

The most important file sits at the very top. I think of it as the read-me-first note, and mine is called CLAUDE.md because Claude’s file-based tools look for that name automatically. It tells the AI who I am, what I am working on, how I like to work and which notes to read before it tackles anything big. That one file is what replaces the spiel. Here is a made-up example of what one could look like:

# Read me first

## About me
Sam, a product manager at a mid-size software company.
Two kids, little free time, prefers direct answers.

## What I'm working on
Q4 launch plan, a side project, a trip to Lisbon.

## How I like to work
Answer first, then explain. Short paragraphs.
Tell me when you're unsure instead of guessing.

## Read these before big tasks
[[My Priorities]], [[Current Projects]]

Beyond that file, the vault holds the notes you would expect: a person’s profile, a project’s status, a decision and why I made it. I also keep notes about my own writing style, so that when I ask AI to draft something, it sounds like me and not like a press release.

A template I recommend: the decision note

The structure of a note matters as much as the folders. I keep about twenty templates, and the one I would suggest to almost anyone is the decision note. The few lines at the top, called properties, let you and your AI filter notes by type, date or outcome. The headings underneath make me think before I commit.

---
type: decision
date: YYYY-MM-DD
outcome: pending
domain: 
---

# Decision: <title>

## Context
What situation prompted this decision?

## Options Considered
1. 
2. 

## The Decision
What was chosen and why?

## Outcome
What actually happened?

## Lesson
What would you do differently? What does this teach?

The payoff comes months later. You can ask your AI, “Look at my decision notes where the outcome was reversed and tell me what pattern you see in my lessons.” It can only answer that question because the notes were written the same way every time.

Three ways to plug your vault into AI

Obsidian does not connect to AI on its own, so here are three ways to do it, from least to most effort.

1. Paste or attach. Copy the notes you need into a chat, or attach the files. It works with any AI and takes no setup. The cost is that you do it by hand every time.
2. Upload to a project. Claude Projects and ChatGPT Projects let you upload files that every chat inside that project can use. It is a good middle step, but it is a copy, so you have to re-upload when your notes change.
3. Give a tool access to the folder (Recommended). Some AI tools, such as Claude’s desktop app and Claude Code, can be given access to a folder on your computer. They read the notes directly and can write back to them. This is how I work, and it is what makes the vault feel alive.

My advice is to start with option 1 if you aren’t really sure and need to test it out first. Move to option 3 once you have enough notes that copying them over starts to annoy you. If you use option 3, tell your AI to read your read-me-first note at the start of every session.

What actually changes: a before and after

Here is a made-up example, using the fictional Sam from the sample note above. Sam types the same request twice: “Help me plan the next two weeks of my launch.”

Without a vault, the AI has only that sentence, so it replies with a generic plan:

Start by defining your launch goals, identifying stakeholders, building a timeline and drafting a communications plan. Here is a template you can adapt…

It is not wrong, but Sam has to answer a dozen follow-up questions before it becomes useful.

With a vault, the AI has already read Sam’s read-me-first note, the launch project note and a decision note from last month. It replies:

Your launch is December 3 and your team is three people, so I’d front-load the week of the 9th for sign-offs. You decided in October to cut the reporting feature, so I’ve left it out of the checklist. You prefer short answers, so here is a one-page plan. Two open questions from your project note still need an owner…

Same model and same request. The difference is the context in the window, and Sam typed none of it.

Why it’s a no-brainer for me

My verdict is that Obsidian is a no-brainer in the age of AI, for both work and personal life. AI can update your vault for you, so it becomes a living repository of thoughts and ideas that are connected, which is just how I see the world.

I also context switch more than ever. While I wait for one chat to answer, I start another with different questions and ideas. Without Obsidian, I would not be able to keep track of all my thoughts and remember where I left off.

A vault also keeps me from locking into any one AI. I love Claude, but I would still like to keep my options open. My notes are plain files that belong to me, so if another tool turns out to be better for a certain job, I can point it at the same vault and it knows what I know.

My vault is also the foundation for the Hermes agents I built to help me run my household finances and create custom enrichment curriculums for my girls. And my everyday experience with AI is much more tailored to me, and therefore more helpful and practical, because my vault holds most of what it needs to know.

Where it falls short

A vault is only as good as what is in it. If your notes are thin, your AI’s answers will be too. If they go out of date, the AI will confidently work from old information, so it is worth skimming what it writes into your vault now and then.

It also takes a little setup, and the habit of writing things down. For me that cost is small next to the time I get back, but it is real.

There is one more thing worth thinking about, which is what you choose to put in your vault and which tools you let read it. I share my own thoughts on that in Where I Draw the Line on AI and Privacy.

How to start in 15 minutes

You do not need my whole system. Start small:

1. Download Obsidian and create a vault, which is just a new folder.
2. Write three short notes: one about who you are and what you do, one about what you are working on right now, and one about how you like to work.
3. Link them to each other with double brackets.
4. Start your next AI chat by attaching or pasting those notes, then ask the question you actually care about.

If you read Part 1 of my personality series, your saved test results make a great fourth note. From there, add to the vault whenever you catch yourself explaining the same thing to AI twice.

Sample prompts to seed your vault

An empty vault is the hardest part, so let the AI do the first draft. These are starting points I wrote for this post. Paste one in, answer honestly, read what comes back, then save it as a note.

1. Interview me. This builds your first note.

I'm building a personal knowledge base in Obsidian so you can know me without me re-explaining myself every time. Interview me one question at a time about who I am, what I do for work, what I'm focused on right now, and how I like to work. Ask follow-ups when my answers are vague. After about 10 questions, write a short "About me" note in markdown. Only use what I told you, and flag anything you're unsure about.

2. Write my read-me-first note. Run this after the interview.

Using what I've told you, draft a read-me-first note under 300 words. Include who I am, what I'm working on, how I like you to respond, and a list of notes you should read before big tasks. Write it in the first person so I can edit it easily.

3. Turn what I already have into notes. Good for a resume, bio or old project doc.

Here is my resume. Turn it into separate markdown notes, one per topic, with [[links]] between related notes. Don't invent details. If something is missing, list your questions at the end.

4. Capture my day. A brain dump becomes a daily note.

Here's a brain dump of my day. Turn it into a daily note with these sections: what happened, decisions I made, open questions, and what I want to remember. Link any people or projects that already have notes.

5. Keep it fresh. Run this once in a while.

Read my "Current Projects" note and flag anything that looks out of date, contradictory or missing. Suggest edits, but don't change anything until I approve.

The common thread in every prompt is telling the AI to stick to what you said and to ask when it is unsure. That keeps the vault true to you, and a vault you can trust is the whole point.

Where to go after week one

Once the basics feel natural, plugins are where Obsidian gets powerful. Plugins are free add-ons you install from inside the app. The ones I use are Templater, which fills in dates and titles when you create a note from a template, Periodic Notes, which creates daily and weekly notes for you, and Dataview, which lets you search your notes like a database. I also use Tasks to track to-dos across notes and Obsidian Git to back up my vault and keep a version history.

My daily note setup works really well for me, and it deserves its own space. I will cover it, along with my plugin setup, in follow-up posts.

Frequently asked questions