Last week, Andrej Karpathy posted something that went viral. Not a new model or a benchmark. A workflow.

Karpathy was a founding member of OpenAI, led AI at Tesla, and coined the term “vibe coding.” What he shared this time was a shift in how he uses AI. In his words, “a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge.”

The system is simple. He dumps raw research material into a folder. Papers, articles, repos, datasets. An LLM compiles it into a structured wiki of markdown files. It writes encyclopedia-style entries, cross-links concepts, and runs health checks to catch inconsistencies. 100 articles and 400,000 words on a single research topic. He didn’t write any of it directly. No RAG pipeline. Just folders, a schema file, and a model that acts like a full-time librarian. “My own explorations and queries always add up in the knowledge base,” he wrote. The system grows from use.

Within days, dozens of tutorials appeared explaining how to replicate his setup. But the tutorials all focus on the same thing. The three folders, the schema file, the compilation step. They miss the bigger idea. The interesting question isn’t how to organize your notes with AI.

It’s what happens when your AI has persistent memory of everything you’ve consumed, everything you care about, and everything you’ve ever asked it.

Here’s the reality of AI in 2026. Every conversation starts from zero.

You open Claude or ChatGPT, rebuild context, get something useful, close the tab. Next morning, it’s gone. You start over. The cost isn’t in dollars. It’s in knowledge that never compounds.

The industry’s answer has been RAG. Retrieval-Augmented Generation. You take your documents, chop them into chunks, convert them into vectors, and store them in a specialized database. When you ask a question, the system retrieves relevant chunks and feeds them to the model.

It works, sort of.

But as Karpathy put it in his follow-up, “the LLM is rediscovering knowledge from scratch on every question. There’s no accumulation.”

It’s a search engine with extra steps.

Karpathy skipped all of that. His argument is simple. At personal scale, a few hundred documents and a few hundred thousand words, the LLM can just read the whole thing. No chunking. No embeddings. No vector database. You compile the knowledge into a structured wiki, keep an index that fits in context, and let the model reason over it directly.

The counter-signal here matters. A lot of venture money has gone into vector databases and RAG infrastructure for personal knowledge management. One of the most respected AI researchers alive is saying that, at personal scale, he doesn’t need any of it. Just three folders and a schema file.

I wrote an essay earlier this year called “You Are What You Consume.” The argument was that in a world of infinite content, the only leverage is what you choose to let in. I described how I treat my inputs like a diet. How I use tools like Goodreads, Beli, and Letterboxd to make my taste explicit. How I built a script to summarize my newsletters each morning and surface only what matters. How I set up AI to monitor the sites I visit and highlight what’s actually relevant.

That essay was about the first step. Curating your inputs. Deciding what gets in.

Curation alone doesn’t compound. Even the good information just sat there. Bookmarked but never revisited. Saved but never connected. The articles I read, the companies I researched, the ideas I had at 11pm. All of it scattered across apps and tabs and notes that I’d never look at again.

Karpathy’s post clarified the second step. Compilation. Taking raw material and giving it structure. Connecting ideas. Building something the AI can reason over.

But there’s a third step that I think matters more than either of those, and it’s the one nobody is writing about. The knowledge base isn’t just a personal wiki for you to browse. It’s the memory layer for an AI that learns who you are.

Every article you save teaches the model what you care about. Every note you write reveals how you think. Every question you ask shapes what connections the model draws next. Over time, you’re not just organizing information. You’re building a persistent representation of your knowledge, your taste, your interests, and your open questions. You’re training an AI that actually knows you.

Curation decides what gets in. Compilation gives it structure. Persistent memory turns it into something that can serve you back.

My version runs on Notion, not Obsidian. I was already living in Notion. My restaurants, my travel recommendations, my reading notes, my investment research, my writing pipeline. Years of accumulated structure. When I started building an AI-augmented knowledge system, starting over in Obsidian with empty markdown files made no sense. The best system is the one you already inhabit.

Obsidian has real advantages for this. Markdown is the most LLM-friendly format, bulk read/write is trivial, and there’s no API layer between the model and the data. For a single research topic, that simplicity is a real asset.

But Notion gives me something flat files don’t. Structured metadata. Every entry in my ideas database has properties. Area, topic, status, importance, source links, next steps. I can filter, sort, and query across hundreds of entries without the LLM having to parse unstructured text to figure out what’s what. When I’m researching across multiple domains simultaneously, investment themes, essay ideas, companies, technologies, that structured layer pays for itself.

The AI connects through Claude’s MCP integration, the Model Context Protocol, which lets the model read from and write to Notion directly. No copy-pasting. No manual syncing. When I’m researching a company or exploring an investment thesis, the model pulls context from my existing notes, cross-references what I’ve already captured, and deposits new findings back into the database with the right properties filled in.

The intake layer feeds it. That newsletter summarization script I built? Every morning it processes the previous day’s newsletters and surfaces the key signals. The best material doesn’t just get summarized and forgotten. It gets routed into the Ideas database with structured tags, source links, and a relevance assessment. Articles I read during the day follow the same path. The raw material accumulates with structure already attached.

What I’m still building is the maintenance loop. Weekly cross-linking passes that scan for connections between entries I wouldn’t notice manually. Cleanup routines that flag incomplete entries and stale references. The goal is the same self-healing quality Karpathy described, applied to a structured database instead of a folder of markdown files. The compilation step works. The autonomous maintenance is the frontier.

The tutorials all stop at “now you have an organized wiki.” That’s useful. But it’s not the end state. When your AI has persistent memory of your knowledge base, it stops being a tool you query and starts becoming a research partner that knows your history.

It can retrieve with real context. Not just “find me the article about X” but “find me what I’ve already read that connects to this new thing I’m looking at.” The model knows what you’ve consumed, so it surfaces your own past research instead of starting from the internet every time.

It can connect ideas across domains. An investment thesis I captured three months ago links to an essay idea I had last week links to a company someone mentioned in a newsletter yesterday. Those connections exist in my database. Without persistent memory, I’d have to notice them myself. With it, the AI draws lines between things I’ve already collected but haven’t consciously connected.

It can recommend based on your actual taste and knowledge. Not “people who liked X also liked Y.” Something more like “based on what you’ve researched, what you’ve written about, and where your open questions are, here’s what you should look at next.” Recommendations grounded in your own knowledge graph, not a popularity algorithm.

And it can act. Not just answer questions but proactively flag when something new contradicts something in your existing worldview. Or when a company you’ve been tracking makes a move that connects to a thesis you wrote about months ago. The knowledge base becomes the foundation for an agent that doesn’t wait.

I’ve already started building this. I have a personal agent that pulls from my knowledge base and searches for relevant new information while I’m sleeping. It’s early, but the loop is real.

Karpathy described his setup as “a hacky collection of scripts” and added, “I think there is room here for an incredible new product.” I built mine by wiring together Notion, Claude, and custom automation over months. The fact that everyone building this today has to do it from scratch tells you everything about the market.

Companies are working on the AI memory layer. Mem0, Letta, Honcho. But nobody has shipped the thing that makes this accessible to someone who can’t wire together APIs and write schema files.

ChatGPT’s memory feature stores fragments. Claude’s memory captures snapshots. Neither builds a structured, queryable, self-maintaining knowledge base that compounds from use. They remember facts. They don’t compound understanding.

The market is still wide open.

If knowledge systems compound, then the gap between people who build them and people who don’t will widen over time. Not linearly. Because the person with a compounding knowledge base doesn’t just have more information. They have more connections. More context. More surface area for new ideas to attach to.

Today, this is a power-user tool. You need to be comfortable with prompts, APIs, file structures, and automation. The barrier to entry is high.

But the same was true of spreadsheets in 1985. And personal websites in 1995. And social media in 2005. The tools get easier. The advantage shifts from “can you build it” to “did you start early enough that your system has years of compounding behind it.”

I wrote before that curation is the foundation. That you’re training yourself whether you’re intentional about it or not. This is the next layer. You’re not just curating what you consume. You’re teaching your AI what you’ve consumed, how you think about it, and what to do with it next. The people who start building that memory layer now won’t just have better notes. They’ll have better agents.

This becomes normal. The only question is who gets it right.

Thanks for reading Mixed Realities by TJ Kawamura! Subscribe for free to receive new posts and support my work.

Subscribe