Build a Knowledge Base Your AI Actually Maintains for You

The Folder of Articles You Swore You'd Read Again

Be honest with me for a second. Somewhere on your computer there is a folder, or a browser full of bookmarks, or a Notion page, stuffed with articles you saved because they were useful for a project. And you have never opened a single one of them again.

I do this constantly. I research a topic for a course, read fifteen good sources, build the thing, ship it, and then six months later a related project lands and I start the whole research process over from scratch. The knowledge I gathered the first time is technically still sitting on my hard drive. It just may as well not exist, because it is not organized in any way I can actually use.

There is a new approach going around that fixes exactly this, and the nice part is that you do not need to be technical to use it. The idea is to stop treating AI like a search box you ask one question at a time, and start using it to build and quietly maintain a knowledge base you can keep asking questions of for years. Let me walk you through it.

Where This Idea Came From

On April 4, 2026, Andrej Karpathy (he was a founding member of OpenAI and led AI at Tesla, so he knows what he is talking about) published a short writeup he called the "LLM Wiki." It got a lot of attention fast, and for good reason. The concept is simple enough to explain in a sentence.

Instead of dumping your sources into an AI every time you have a question and making it re-read everything from zero, you have the AI read your sources once and write them up into an organized, interlinked set of notes. Then you ask your questions against the notes. The notes stick around. They get better every time you use them. The AI does the boring upkeep, and you do the thinking.

Karpathy put the division of labor really cleanly: the human's job is to pick good sources, steer the research, and ask good questions. The AI's job is everything else. If you are an instructional designer, that split should sound extremely familiar, because it is basically how you already work with a subject matter expert.

What "Knowledge Base" Actually Means Here

Let me strip the jargon out, because the words make this sound more complicated than it is.

A knowledge base, in this approach, is just a folder of plain text notes. That is it. No database, no special software, no code. The notes are written in something called Markdown, which is a fancy word for "text with a few simple formatting marks like # for a heading." If you have ever typed a bullet point with a dash, you already understand Markdown.

The setup has three simple parts:

• Your raw sources. A folder where you drop everything you collect: articles, PDFs, meeting notes, slide decks, reports. You never edit this stuff. It is your evidence pile.
• The wiki. A second folder where the AI writes its organized version of everything. It reads your sources and turns them into clean topic pages, with links connecting related ideas together. You rarely touch this. It belongs to the AI.
• The instructions. A short note telling the AI how you want things organized and what to do when two sources disagree.

Then there are three things you do with it: feed it new sources, ask it questions, and every so often ask it to clean itself up. That last one is the part that genuinely surprised me, so let me come back to it.

Why This Is Such a Good Fit for L&D

Most AI advice for our field is about generating content faster. This is different, and honestly more useful, because it goes after problems we have had for decades. Here are the ones it actually solves.

The research pile that never pays off

This is the one I opened with. When you build a knowledge base instead of a bookmark graveyard, your research compounds. The reading you did for the compliance course feeds the onboarding refresh. The sources you gathered on adult learning theory are still organized and queryable a year later. Nothing you read is wasted, because it all lands in one place that gets smarter over time.

Wikis and SharePoint sites that rot

Every L&D team has built a shared knowledge hub that everybody loved for about three weeks and then quietly abandoned. It did not fail because people are lazy. It failed because keeping a wiki current is a grind, and the busywork piles up faster than anyone can keep up with it. The whole point of this approach is that the AI does the grind. It does not get bored, it does not forget to update a cross-reference, and it can update ten to fifteen linked pages in one pass. The maintenance problem that killed every wiki you ever built is the exact thing the AI is good at.

Capturing what's in a SME's head

Getting knowledge out of a subject matter expert and into something usable is slow, awkward, and fragile. When that person leaves, a lot of it walks out the door with them. With this setup, you can drop the raw stuff a SME already produces (emails where they explained something, a recorded call, their old slide decks, their scribbled SOPs) straight into your sources folder. The AI reads it, organizes it by topic instead of by document, and flags where two things they said contradict each other. You become the editor and quality check instead of the transcriber.

A nod to the forgetting curve

One thing I find genuinely interesting. We all know Ebbinghaus and the forgetting curve. People forget most of what they learn within days unless they revisit it. Human memory decays without reinforcement. This kind of knowledge base does the opposite. Every time you add a source or ask a question, it gets a little stronger and a little more complete. It is built to compound instead of decay, which is a nice thing to have around when your own brain is doing the normal human thing and leaking information.

How a Total Beginner Starts This Week

Here is the part I really want you to take away, because you can do this without installing anything fancy or writing a line of code. Start small.

Phase one: just your browser

You do not need any special tools to feel the difference. Try this with a project you are working on right now:

1. Pick one topic you are actively researching.
2. Gather eight to ten good sources. Articles, a couple of PDFs, your own notes, whatever you have.
3. Open whatever AI chat tool you already use. Paste the sources in.
4. Ask it to read all of them and write up an organized set of notes: a short summary of each source, topic pages for the main ideas, and links noting how the ideas connect.
5. Save what it gives you somewhere. Even a Word doc is fine to start.
6. Now ask it real questions against those notes. "Based on these, what are the five things a new hire absolutely has to understand?" "Write learning objectives for each." "Draft a three-module outline."

The first time you do this, you will notice the answers are sharper and more specific than what you usually get from AI. That is because the AI is working from sources you chose, not from a vague memory of the whole internet. That alone is worth the half hour it takes.

Phase two: add a real home for it

Once you are sold, give your knowledge base a proper place to live. The tool most people use for this is Obsidian. It is free, it runs on your own computer, and at its core it is just a nice way to view a folder of notes. No coding, no account required to get going.

There is also a free browser add-on called the Obsidian Web Clipper that saves any web article straight into your notes folder as clean text with one click. That removes the most annoying part of all this, which is the copying and pasting. You clip as you research, and your sources folder fills itself.

Obsidian also has a graph view that draws every note as a dot with lines showing how they connect. The first time you see your own research laid out as a little web of linked ideas, it clicks in a way that a folder of files never does.

Phase three: let it clean up after itself

This is the move that turns a pile of notes into something that feels alive. Every so often, ask the AI to do a health check on your knowledge base. "Look through all of these notes. Where do two of them contradict each other? What topics did I start and never finish? What pages have nothing linking to them? What connections am I missing?"

It comes back with a genuinely useful list. It catches the policy note that conflicts with the procedure note. It finds the orphan page you forgot about. It suggests new topics worth pulling in. This is work no human ever has the patience to do across a growing pile of notes, and the AI just does it. Over time your knowledge base gets cleaner instead of messier, which is the exact opposite of how every shared drive you have ever used behaves.

The Honest Caveats

I am not going to pretend this is magic, so here is the straight talk.

You still have to curate. The AI organizes whatever you give it. If you feed it junk sources, you get a beautifully organized pile of junk. Picking what is trustworthy and what actually matters for your learners is your job, and it always will be. That is the instructional design skill no tool replaces.

Check the important claims. AI can get things wrong or smooth over a distinction that matters. For anything high stakes, especially compliance, safety, or regulated content, you read it and you verify it against the original source before it goes anywhere near a learner. Treat the AI's notes as a smart first draft, not gospel.

It works best at a human scale. This approach shines for a personal or small-team knowledge base, the kind built from a hundred or so good sources. It is not an enterprise content system for your entire LMS. Keep your expectations matched to that and you will be happy.

Why I Think This Is Worth Your Time

Most of the AI conversation in our field is about speed. Make the deck faster, write the quiz faster, draft the script faster. That is fine, but it is also a little shallow, because faster output of stuff you then throw away is not actually progress.

This is the first AI workflow I have seen that makes your work add up. Every project you research leaves you with something permanent and reusable instead of a folder you abandon. For a beginner, that is honestly a better place to start than yet another "write me a course outline" prompt, because it builds a habit and an asset at the same time.

So this week, pick one project, gather a handful of sources, and ask your AI to organize them into notes you can question. Half an hour. That is the whole ask. The version of you starting the next related project will be very glad you did.