An LLM wiki is a set of pages an AI writes and maintains about your source material, with links between related pages and an index at the top. The AI reads the index first and then only the pages it needs. The idea is usually credited to Andrej Karpathy.

How do I make Claude update my knowledge base automatically?

In this video, Claude Cowork is pointed at the notes folder and given a scheduled task that runs every night at 2 a.m. The task checks for new files and adds them to the index and wiki pages. Keep the folder under version control so you can undo a bad run.

Do I need Obsidian for an AI knowledge base?

No. Obsidian is only a viewer for a folder of markdown files, and the AI works on the folder directly. It is useful because it shows links between pages and offers the Web Clipper extension, but any editor or file browser can open the same files.

All video notes

Video notes · Notes and PKM

Build a Self Improving Claude Knowledge Base

Video by Paul J Lipsky on YouTube · 14 min · uploaded June 2026

Notes and analysis by BrainTube, written from the video. Not affiliated with Paul J Lipsky. Watch on YouTube

In short

The video shows how to turn a folder of notes into an LLM wiki that Claude Cowork reorganises and indexes every night, viewed through Obsidian. The core idea, a maintained index the AI reads instead of rescanning every file, is sound and comes from Andrej Karpathy. The how is mostly hidden in a pasted prompt, and the video says little about cost, errors or what "self-improving" really means.

The argument

Lipsky's claim is that most people's AI setups are static: you point the assistant at files and it has to work out what they are each time, so it never gets faster. His fix is a map. The AI sorts sources into folders, writes wiki pages that connect related material, and keeps an index that acts as a table of contents for itself. A scheduled task runs every night, checks the folder for anything new and folds it into the map. Because the connections already exist, later questions can be answered from the index and the linked pages instead of from raw files. He applies the same pattern to his YouTube work, a business journal and a health log, and argues it works for any area of life.

Key ideas, with timestamps

0:40
An LLM wiki as a map
He credits the idea to Andrej Karpathy's LLM wiki and calls it a map for the AI. Without one, every new session rereads the folder to work out how files relate, and that work is thrown away a week later. With one, the AI reads the map first and goes straight to what it needs.
2:02
What his knowledge folder holds
His folder has sections for people, sponsors, tools he has covered and topics. Most of it comes from transcripts of his own videos, plus saved articles, tweets and other creators' videos. Opening the Gemini page at 2:44 shows his videos and saved articles about Gemini grouped with notes on what matters about it to him.
3:25
A second wiki for health
He runs a separate knowledge base for health that logs sleep, calories, weight, resting heart rate and HRV into daily pages and produces a readiness score each morning. He then asks it how he is tracking against his goals. It is the clearest example in the video of the pattern working outside content.
4:45
Obsidian as the viewer
Obsidian is used only to look at the folder. He explains that a vault is just a folder, drags existing notes in, and at 6:08 adds the free Obsidian Web Clipper extension so articles and YouTube pages can be saved straight into the vault.
6:08
Pointing Claude Cowork at the folder
In Claude Cowork he selects the vault folder so it can read and write the files, then shows the connectors panel for email, calendar and Notion. From 7:31 to about 9:10 a sponsored segment explains connecting Zapier's MCP server for apps without a native connector, with per-app permissions; he uses it to pull from his YouTube account.
9:33
One prompt builds the system
The actual build is a prompt from a linked Google Doc, pasted into Cowork with the folder selected. Cowork proposes a folder plan and asks for approval, asks where the index should live (he picks the recommended wiki folder), then builds wiki pages, a sources folder, a review file and the index.
10:55
The nightly 2 a.m. task
The last step is a scheduled task that runs at 2:00 a.m. every day, checks for new files and adds them to the map. This scheduled pass is what he means by self-improving: the structure is refreshed without him asking.
11:36
Asking for patterns
Once built, he says the AI uses the wiki on its own, for example noticing that a new client problem resembles an old one. He also asks direct questions such as which patterns show up in his thinking or which revenue streams he is missing. At 12:16 he adds a business journal that links each session to earlier entries.

What it gets right

The index-first idea is the right instinct. An assistant that reads a short table of contents and then opens three linked pages is cheaper and more focused than one that skims a hundred raw files every session. Separating raw sources from the AI's own wiki pages is also good practice, because the originals stay untouched and the summaries can be rebuilt.

Keeping it in a plain folder viewed through Obsidian means nothing is locked in. If the setup stops working you still own readable markdown. The health example is a fair test of the pattern, since daily logs with a few numeric fields are exactly what an LLM can tidy and summarise well.

What it leaves out

The prompt is the method. Everything that decides quality, such as how pages are split, how links are chosen and what the index records, sits in a Google Doc the video never walks through. Viewers learn the clicks, not the design.
"Self-improving" is a scheduled rewrite. The nightly job reorganises and links files. It does not check whether its summaries are right. A wrong connection made at 2 a.m. becomes part of the map the next session trusts. The review file is there, but he never shows reading it.
Unattended writes. At 10:13 he says Cowork never acts without permission, yet the 2 a.m. task runs while he sleeps. Keep the vault under version control so a bad night can be rolled back.
Scale and cost. One index file works for hundreds of notes. With thousands, the index itself gets long, and a nightly pass over everything costs tokens every day. Searching sources directly, by keyword or meaning, is often the better tool for lookups.
Sensitive data. Heart rate, weight and business journals are sent to the model provider on every pass. That deserves a sentence, and it gets none.
Results. The intro claims breakthroughs in his business and health. The video shows a readiness score and a list of folders, not an outcome.

How this applies to memory across AI tools

The wiki lives in a folder that Claude Cowork can open. Other assistants only see it if they also get file access, or if something serves it to them over MCP. That is why the Zapier segment matters more than the sponsor read suggests: MCP is the route by which a knowledge base becomes reachable from more than one tool. Read persistent memory vs the context window for why a map beats rereading files.

If most of your sources are videos, articles and PDFs, you can skip clipping them into a vault. BrainTube saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes it as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search it. Personal logs like the health pages still fit well in a local wiki. See how to remember what you watch and read for the source side.

Who should watch it

Watch it if you use Claude Cowork and want a quick look at what a maintained LLM wiki looks like in Obsidian, or ideas for applying one to health or journaling. The demo of the finished wiki is at 2:02 to 4:05, and the build itself is 9:33 to 11:36. Skip 7:31 to about 9:10 unless you need Zapier, since it is a sponsored segment. Skip the video if you want to understand how the wiki is structured or why it works; that lives in the linked prompt, so read the prompt instead.

Frequently asked

What is an LLM wiki?: An LLM wiki is a set of pages an AI writes and maintains about your source material, with links between related pages and an index at the top. The AI reads the index first and then only the pages it needs. The idea is usually credited to Andrej Karpathy.
How do I make Claude update my knowledge base automatically?: In this video, Claude Cowork is pointed at the notes folder and given a scheduled task that runs every night at 2 a.m. The task checks for new files and adds them to the index and wiki pages. Keep the folder under version control so you can undo a bad run.
Do I need Obsidian for an AI knowledge base?: No. Obsidian is only a viewer for a folder of markdown files, and the AI works on the folder directly. It is useful because it shows links between pages and offers the Web Clipper extension, but any editor or file browser can open the same files.

Keep videos like this one searchable

BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.

Start free

Build a Self Improving Claude Knowledge Base

In short

The argument

Key ideas, with timestamps

An LLM wiki as a map

What his knowledge folder holds

A second wiki for health

Obsidian as the viewer

Pointing Claude Cowork at the folder

One prompt builds the system

The nightly 2 a.m. task

Asking for patterns