Why AI Keeps Forgetting Everything You Tell It

There’s a moment everyone hits with AI.

You upload a stack of documents — research papers, meeting notes, articles you’ve saved. You ask a question. The answer is good. Really good. It pulls from the right places, connects things you hadn’t connected yourself. You think: this is it. This is the tool.

Then you come back the next day. Ask a similar question. The connections are gone. The synthesis it did yesterday isn’t there. The AI can read the same files again, but the work it already did has evaporated.

This isn’t a bug. It’s how the whole thing works. And the half-measures most people use to fix it — ChatGPT Memory, Projects, Custom GPTs, NotebookLM — solve a slice of the problem without solving the core one.

What the half-measures actually do

ChatGPT Memory remembers facts about you. Your name, your preferences, that you have three kids. It doesn’t read your documents and synthesize them.

Projects and NotebookLM keep a fixed set of files in scope, so the AI has them at hand each time. That helps. But the AI is still doing the same job from scratch every session — searching the files, grabbing chunks, generating an answer. The synthesis doesn’t persist between sessions. Whatever connections it found yesterday are gone tomorrow.

So the question isn’t “how do I make the AI remember me.” It’s “how do I make the AI’s thinking about my material persist.”

The filing cabinet problem

Here’s what’s happening under the hood, in plain terms.

When you point an AI at a folder of documents, it doesn’t absorb the material the way you would if you sat down and read it. It searches through the files when you ask a question, grabs whatever chunks look relevant, and generates an answer from those pieces. This is called retrieval. It’s the standard pattern.

Think of it like having an assistant with a massive filing cabinet. Every time you ask a question, they open the cabinet, rifle through some folders, grab a few pages, and write you a response. Then they put everything back, close the cabinet, and forget they ever opened it. Tomorrow, same thing. Same rifling. Same starting from scratch.

For simple questions — “what did the Q3 report say about revenue?” — this works fine. The assistant finds the page, reads the number, gives it to you.

But what if your question requires connecting ideas across five different documents? What if the answer lives in the relationship between things, not in any single page? The assistant has to find all those pieces, hold them together, and synthesize something useful. Every single time. With no memory of having done it before.

The synthesis doesn’t compound. Every question is a cold start.

What if the AI actually read the material?

Andrej Karpathy recently shared a workflow that reframes the whole thing. He didn’t invent every piece of it — there’s a long lineage here, from Zettelkasten to second brains to RAG-with-summarization. But he packaged it cleanly enough that a lot of people, including me, looked at it and went: oh.

The idea: instead of searching raw files every time, you have the AI read your documents once and build a structured knowledge base out of them. Organized pages, linked concepts, summaries, comparisons — all maintained by the AI itself.

When you add a new source — a PDF, an article, a set of notes — the AI doesn’t just store it for later. It reads it, extracts the key ideas, and integrates them into the knowledge base. It updates existing pages. Creates new ones for new concepts. Links related ideas together. If the new source contradicts something already in the knowledge base, it flags that.

So over time, the thing grows. The connections are already there. The synthesis is already done. When you ask a question, the AI isn’t starting from scratch — it’s working from a pre-built body of knowledge that gets richer with every document you feed it.

Karpathy calls it the LLM Wiki. The name is sticky. The pattern is older than the name.

How it actually works (the part most posts skip)

The runtime matters. This is where most explanations of this idea wave their hands and lose you.

You need an AI agent that can read files in a folder, write files in another folder, and do that on its own when you tell it to. The consumer ChatGPT app can’t do this. Neither can the Claude web app. You need a tool that gives the AI access to your file system.

The options today:

Claude Code — Anthropic’s CLI tool. Drops into a folder, reads and writes files, runs the agent loop. What I use.
Cursor agents or other agentic IDE tools — same idea, different wrapper.
A custom script — anyone comfortable with the API can build the loop in an afternoon.
NotebookLM if you want the closest no-setup option, though you give up the file-on-your-disk part.

Once you have the runtime, the structure is simple:

Raw sources. A folder of original documents — PDFs, articles, notes, whatever you’re working with. Read-only. The AI reads but never changes them. This is your source of truth.

The wiki. A folder of organized pages the AI creates and maintains. An index page. Concept pages. Entity pages. Summaries and comparisons. All interlinked, all updated by the AI as new sources come in.

The rules. A short document that tells the AI how to operate — how to structure pages, how to handle new sources, how to format links. Think of it as the operating manual you hand to a new research assistant on their first day.

That’s it. Three folders, a rules document, and an agent that can read and write files.

What it looks like in practice

I’ve been running a version of this for the last few months on a Brain folder. Reading list, investor research, building notes, articles I want to remember.

You drop a new article into the raw folder. Tell the agent to read it and update the wiki. It reads the article, extracts the key concepts, creates structured pages for each one, links them together, and updates the index. A few minutes later, you have organized, interlinked pages covering the article.

Then you add a second source. A different article on a related topic. The agent doesn’t just create new pages — it goes back and updates the existing ones. It finds connections between the two sources. New links form. Existing pages get richer.

When you ask a question that spans both sources, the agent pulls from the wiki, not the raw articles. It connects dots that were spread across completely different documents and cites its sources.

The first time you watch this happen with your own material, it lands.

Linting your knowledge

Here’s a detail I think is clever. Just like software developers run tools to check their code for problems, you can periodically ask the agent to audit the wiki. Look for contradictions between pages. Claims that might be outdated. Pages with no connections to anything else. Concepts that get mentioned but don’t have their own page yet.

You’re maintaining the knowledge base the way you’d maintain a codebase. Pruning, connecting, resolving conflicts. Keeping it healthy as it grows.

What compounds and what doesn’t

I want to be careful here, because I caught myself reaching for an investor analogy that doesn’t quite hold.

The honest version: the AI’s synthesis compounds. The work it does to connect ideas across documents persists, instead of evaporating. So the second time you ask a hard cross-document question, the answer is faster and better, because the connections were already drawn.

What doesn’t compound: the audit cost. Wiki rot is real. Pages go stale. The AI sometimes draws a wrong connection or summarizes a source in a way that drifts from the original. The longer the wiki runs, the more you have to maintain it. That’s why the audit step matters.

The right way to think about it isn’t “compound interest.” It’s closer to a garden. The work you put in stays there, but you have to weed.

The honest limitations

This isn’t magic, and a few things will bite you:

Cost. Re-reading and rewriting wiki pages with frontier models adds up. Not crippling at personal scale, but not free either. Budget for it.

Hallucinated links. The AI will sometimes connect two concepts that look related but aren’t. Spot-check links during audits, especially for anything you’d cite or act on.

Versioning. When the AI rewrites a page and gets it wrong, you want to roll back. Keep the wiki in git, or at least back it up.

Hybrid edits. When you edit a page yourself and then the AI re-edits it, the AI can clobber your edits. Decide upfront who has final say on each page type.

Privacy. “Your data stays on your computer” is only true if you’re running a local model. If ChatGPT or Claude is doing the synthesis, your data is going to that provider. Read their terms.

Scale. The pattern works best at personal scale — Karpathy talks about wikis with around a hundred source documents. Past that, you start needing real infrastructure.

With all of that said, this is one of the most practical AI workflows I’ve come across. It solves a real problem. And unlike almost everything else in AI, the synthesis you build keeps building, instead of resetting to zero.

The shift

Most people use AI as a question-answering machine. You go to it when you need something, get your answer, leave. The AI is a tool you pick up and put down. It never gets better at helping you specifically, because the work it does for you doesn’t persist.

The wiki flips that. The AI becomes something closer to a long-term research partner — one that builds on previous work and gets more useful the more you use it. Not because the AI itself is getting smarter, but because the body of knowledge it works from keeps growing.

That’s a different thing entirely.

What to do this week

If you want to try it, here’s the smallest possible version:

Pick a topic you actually care about. Investing, a business problem, a research interest.
Make three folders: sources/, wiki/, rules.md.
Drop five things you’ve already read into sources/. PDFs, articles, notes — whatever.
Open Claude Code, Cursor, or your tool of choice in the parent folder.
Tell the agent to read the rules file, read the sources, and build out the wiki.
Add a sixth source and watch what happens.

The whole thing takes an afternoon. If it clicks, you’ll know within a week.

I’ve been wrong about a lot of AI workflows. This one I keep coming back to.