Karpathy’s Second Brain System Went Viral. Building It with Codex and Obsidian

A few weeks ago, Andrej Karpathy shared how he uses a large language model to organize his personal knowledge base. No new model, no product launch, just a method for letting AI sort and maintain local files. The post pulled over 21 million views.

He later turned it into a Markdown document called LLM Wiki and posted it on GitHub. At last check, it had around 42,000 stars.

What caught my attention was not the popularity. It was the idea underneath: instead of asking AI to search through your materials every time from scratch, you let it continuously organize, connect, and update your knowledge into something that grows on its own.

So I rebuilt the whole thing with Codex and Obsidian.

Why Codex + Obsidian?

Obsidian is a local Markdown-based knowledge management tool with millions of users. Its core appeal is that your data stays on your machine, and bidirectional links combined with a knowledge graph connect scattered notes into a personal network.

Codex is OpenAI’s AI Agent tool, now with millions of weekly active users. It has expanded from a code assistant into a general-purpose agent: it can read and modify local files, execute scripts, and handle multi-step tasks.

The setup is straightforward: Obsidian handles storage and display, Codex handles understanding, organization, and maintenance, and Markdown files are the actual data layer.

Everything in Obsidian is a local Markdown file. A knowledge base is just a folder. Other editors, scripts, and AI agents can all read these files directly. Obsidian syncs automatically when external changes happen.

Codex reads the directory, modifies content, runs scripts, and completes complex tasks. It also reads an AGENTS.md file in the project, so you can define your knowledge base’s folder structure, naming rules, reference patterns, and maintenance workflows in advance.

Think of it this way: Obsidian is the backend, Codex is the AI knowledge manager, and Markdown files are your actual data assets.

Setting Up the Knowledge Base

I created a folder called Knowledge, opened it in both Obsidian and Codex, and handed Codex Karpathy’s LLM Wiki document as the blueprint.

The final structure looked like this:

• raw/: original materials such as PDFs, articles, web pages, and notes
• wiki/: AI-generated summaries, topics, concepts, and conclusions extracted from source materials
• AGENTS.md: rules defining how AI should organize knowledge
• scripts/: scripts for searching, checking for gaps, and detecting broken links
• README.md: usage guide for the knowledge base

Once the setup was done, daily use came down to three actions.

Step 1: Ingest Materials

When I come across a valuable article, I send the link or file to Codex and ask it to add the material to the knowledge base.

Codex asks before doing much on first contact. When it receives a new source, it typically checks which folder to use and whether to extract key points. If you prefer a hands-off approach, you can tell it to save study materials directly without asking. It will write that preference into its operating rules.

One thing I actually noticed after running this for a while. One article mentioned an open-source project with 36 Skills at the time of writing. When Codex processed it, the project had already grown to include more. Instead of just storing the text, Codex checked the project’s actual page and updated the information accordingly.

The result is not one long summary per article. Each piece of content gets broken into individual knowledge points, each connected to related entries already in the base. Knowledge grows around real questions, not around how many articles you have bookmarked.

Step 2: Ask Questions Based on the Knowledge Base

After ingesting materials, you can query the knowledge base directly.

For example, based on the articles I processed, I asked: how do you tell whether a product idea actually has users?

Codex started with a concise judgment framework drawn from the knowledge base, then pointed to the specific pages backing each point.

If you need only the quick answer, the current response is enough. If you want to dig deeper, open the relevant Markdown files for the full context, sources, and reasoning.

The practical advantage of this setup: answers stay concise, but the evidence behind them remains traceable.

Step 3: Run Periodic Lint Audits

The longer a knowledge base runs, the more it accumulates duplicate pages, orphaned notes, broken links, outdated conclusions, and contradictory viewpoints.

I run a full lint audit on a regular schedule. The prompt I use:

Run a lint audit across the entire knowledge base. Check for duplicate pages, orphaned pages, broken links, outdated information, conclusions without sources, and contradictory viewpoints. Fix formatting and link issues directly. For deletions, merges, and viewpoint changes, generate a report first and do not execute automatically.

A knowledge base that only adds content is not particularly useful. It needs to delete, merge, and correct old knowledge on a schedule. Otherwise, AI just helps you produce new information garbage faster.

How This Differs from NotebookLM

I have used NotebookLM a fair amount. It is a solid AI knowledge tool: import PDFs, web pages, YouTube videos, audio files, and Google Docs, then query them with cited answers, generate summaries and audio overviews.

But the two tools solve different problems in practice.

NotebookLM is an excellent reading room. A local knowledge base built with Codex and Obsidian is closer to a programmable knowledge operating system.

Use NotebookLM when you need to quickly read a batch of sources, get Q&A around specific texts, generate summaries for temporary research, or work with sources without setting up anything locally.

Use Codex and Obsidian when you are doing long-term research in one area, want AI to update your own judgments over time, need AI to modify your knowledge structure directly, or want to build secondary products on top of your knowledge base.

The critical difference: NotebookLM works with copies of your sources and does not modify the originals. Codex works directly with open Markdown files it can read, update, merge, and restructure. It can also write code and generate web pages and tools from the knowledge it manages.

My own workflow: NotebookLM for quick source reading, Codex and Obsidian for knowledge worth maintaining long-term.

Taking It Further

Most people who set up Obsidian end up with a knowledge graph that looks impressive in screenshots but does not change how they actually work.

Here is a question worth asking: since these notes are all structured Markdown files, can Codex read through them and generate something you can open in a browser?

I asked Codex to read the entire knowledge base and produce a local HTML dashboard. Here is the request I made:

Read the Markdown files in the current knowledge base and generate a browser-ready HTML knowledge dashboard. It needs: global search, a topic word cloud, tag-based theme distribution, a timeline of knowledge additions and edits, recently updated pages, orphan page and broken link alerts, most frequently mentioned people, products, and concepts, an interactive knowledge relationship graph, and each page must link back to its original Markdown file. Data is read directly from local files with no remote database dependency.

Now Obsidian is not the only way to view the knowledge base. It is the backend. The HTML file is a frontend designed for a specific use case.

The same knowledge base can serve multiple frontends: full detail for personal use, a project board for your team, a filtered public page for your audience.

After running this setup, one thing became clear: a knowledge base should not just be a warehouse waiting to be queried. It should be a data layer that continuously generates interfaces, reports, and products.

Historically, building a knowledge base solved one problem: where do I store my materials?

The more interesting problem now is: can my knowledge be continuously updated, recombined, and turned into actual output?

Karpathy’s LLM Wiki is a solid starting point. Its real value goes beyond letting AI automatically organize your Obsidian notes. It is about turning personal knowledge into something you can program, compute, and display.

In this system: Obsidian is the data backend, Codex is the knowledge manager and developer, Markdown files are long-term stored knowledge assets, HTML and reports and charts are different faces of the same knowledge, and you are the one choosing materials, asking questions, and making final judgments.

A second brain is not a better bookmark manager. It is a system that grows with your work and helps you think, judge, and create.