Do you have these problems when using Claude Code?
・It's a hassle to re-explain the same thing to the AI every time
・It doesn't remember yesterday's conversation
・What you researched gets reset in the next session
・Articles and notes you read in the past eventually disappear somewhere
All of this is caused by the fact that "AI has no memory."
An article that breaks down the "AI External Brain" construction method proposed by Andrej Karpathy—former OpenAI and former Tesla AI lead—into a level that can actually be operated with Claude Code is currently going viral overseas with over 2,100 likes 😳
It was written by @hooeem, a creator who regularly posts viral articles in the overseas AI developer community. This time, it's summarized as a 3-stage guide for everyone from complete beginners to developers.
I will now break down and explain the content in an easy-to-understand way 👇
Original post here: https://x.com/hooeem/status/2041196025906418094
■ Why the current way of using AI is "wrong" in the first place
The original article begins like this:
"Most people use AI as an 'amnesiac search engine.'"
Ask a question → get an answer → close the tab. Start over from scratch the next day. Nothing accumulates. Nothing compounds. You keep burning tokens just to rediscover the same context.
Karpathy's system completely flips this.
- Gather materials. Articles, papers, YouTube transcripts, PDFs, anything related to the topic of interest.
- The AI reads it all and writes a structured Wiki. Summaries, conceptual explanations, connections between ideas, and a master index.
- Ask questions against that Wiki. The AI cross-searches its own accumulated knowledge and returns an integrated answer with citations.
- Answers are automatically saved to the Wiki. The next question benefits from all past work.
- The AI periodically performs health checks on the Wiki. It finds and fixes contradictions, gaps, and outdated information.
The result? A personal knowledge base that gets smarter every time you use it.
If you keep adding information for a month, you'll have a deeply linked knowledge asset that Google Search could never replicate. This is because it's not just an "index," but something "integrated."
According to the original article, this can be used for any theme: crypto markets, medical research, legal precedents, competitor analysis, academic research, philosophy. Anything where you want to accumulate and connect knowledge over time.
■ Level 1: For Complete Beginners (Obsidian + Claude Chat)
No technical skills required. You only need two things:
・Obsidian (Free) ── Download from obsidian.md
・Claude subscription ($20/month Pro, or your favorite AI chatbot)
That's it.
Step 1: Create a Vault (2 mins)
Open Obsidian and click "Create new vault." Just give it a name and choose a save location. A Vault is just a folder. Markdown files inside it are automatically displayed as notes.
Step 2: Create two folders (1 min)
raw ── Folder for raw materials (articles, notes, anything)
wiki ── Folder for knowledge summarized by AI
This is the entire basic structure.
Step 3: Add your first materials (5 mins)
Choose one theme you're truly interested in. Find 3-5 good articles on that theme. Create a note for each in the raw folder and copy-paste the text. Write Source: [URL] at the top.
Don't worry about formatting. The important thing is to get the text in.
Step 4: Let the AI create the Wiki (5 mins)
Open Claude (claude.ai) and use the prompt introduced in the original article. Paste your materials and instruct it to "write a summary of each source, list key concepts, and create a master index."
Claude will return structured output, so save each as a note in the wiki folder.
Step 5: See the magic
Open Obsidian's Graph View (Ctrl+G), and you'll see notes as dots with Wiki links connecting them. This is your knowledge base network.
From here, make it a daily habit: when you find a new article, put it in raw and ask Claude to "process the new source based on the existing index." If there's a contradiction, Claude will flag it with a ⚠️.
In short, Level 1 works with just copy-pasting. No terminal or coding required.
■ Level 2: Full System (3-Layer Architecture + CLAUDE.md)
While Level 1 is "copy-paste operation," Level 2 is a "system where AI creates and manages files itself."
The architecture proposed in the original article is a 3-layer structure.
Layer 1: raw/ (Raw Materials) ── The only source of truth. AI reads this but does not rewrite it. Contains articles, papers, repo docs, datasets, and images.
Layer 2: wiki/ (Compiled Wiki) ── Generated and maintained by AI. Contains summaries, concept articles, person/organization pages, cross-links, indexes, and query outputs. Humans generally don't edit this directly.
Layer 3: CLAUDE.md (Schema) ── A configuration file that teaches the AI the "structure, naming conventions, and executable operations" of this Wiki. Place it in the root of the Vault.
And four operational cycles keep running:
・Ingest ── Import new materials. AI automatically generates summaries, concept pages, and connections.
・Compile ── Build and update Wiki pages. Maintain indexes and integrate new information into the existing structure.
・Query ── Ask questions. AI cross-searches the Wiki and returns answers with citations. Answers are saved to the Wiki.
・Lint ── Health check. Automatically find and fix contradictions, gaps, broken links, and outdated information.
The folder structure shown in the original article is like this 👇
my-knowledge-base/
├── raw/
│ ├── articles/
│ ├── papers/
│ ├── repos/
│ ├── datasets/
│ └── assets/
├── wiki/
│ ├── index.md
│ ├── log.md
│ ├── concepts/
│ ├── entities/
│ ├── sources/
│ ├── syntheses/
│ ├── outputs/
│ └── attachments/
├── templates/
└── CLAUDE.md
All filenames are kebab-case (lowercase hyphen-separated). For example, active-inference.md ✓, Active Inference.md ✗. Source summaries follow the author-year-short-title.md format (e.g., friston-2010-free-energy.md).
CLAUDE.md describes the Wiki structure, naming conventions, specific procedures for each operation (Ingest/Query/Lint), page creation thresholds (concepts appearing in 2+ sources get a full page, 1 source gets a stub), and quality standards (summaries 200-500 words, concept articles 500-1500 words).
The original article says, "Keep this file under 80 lines. Every line eats up the context window."
■ What should you put in your knowledge base?
The original article asks:
"Think about everything you consumed in the last year that just vanished."
・Books you finished and forgot
・Podcasts that changed your thinking
・Articles you saved at 11 PM and never opened again
・Late-night YouTube rabbit holes that taught you more than any course
・Kindle highlights you highlighted and never looked at again
・Research you did before a big decision
・Notes from old projects
・Lessons learned from things that didn't work out
All of it is sleeping somewhere doing nothing. All of this belongs in the Vault.
What if you have no materials? Open a Claude chat and talk for 20 minutes. About work, goals, what you're making now, what you're thinking about. Save that conversation as a Memory file. Just that will make you feel like "Claude knows me" from the first session.
The Vault doesn't have to be perfect to be useful. The important thing is that it's "real."
■ Level 3: Automation (5 Stages)
This is for power users. The original article explains automation in 5 stages.
Level 3-1: One-shot execution via CLI
Open Claude Code in the terminal and have it process all unprocessed files in raw/ with a single command.
Level 3-2: Slash commands
Place Markdown files in .claude/commands/ to use custom commands like /wiki-compile. Turn repetitive workflows into single commands.
Level 3-3: Scheduled execution
Use Claude Desktop's /schedule feature or cron to automatically process new files in raw/ every morning. Clip an article before bed, and the Wiki is updated when you wake up.
Level 3-4: GitHub Actions
Turn your Vault into a GitHub repo. When you push to raw/, Claude Code compiles the Wiki on GitHub Actions. It works even when your PC is off.
Level 3-5: Agent Skills
Place skill files in .claude/skills/, and Claude will automatically detect context and perform appropriate operations. If you say "I put a new file in raw/," Claude will automatically run the Ingest cycle without you typing a command.
Advice from the original article: Start from Level 3-1 and build up as you get used to it. Starting anywhere won't break previous levels.
Furthermore, the community has already built plugins. If you install the wiki-skills plugin, you can use /wiki-init /wiki-ingest /wiki-query /wiki-lint commands immediately. You don't even need to write config files manually.
■ Why "maintenance" was the biggest bottleneck
This is the deepest insight in the original article.
Notion, Evernote, Roam Research... there have been many tools claiming to be a "second brain." But most people stop using them after a few months.
The reason is the same: maintenance is too much trouble.
As the original article says: "Putting information in is fun. But organizing tags, updating cross-references, reorganizing structures—when this extra work piles up, it becomes work on top of your actual work. If you slack off, the system degrades. Six months later you try to rebuild it, and the cycle repeats."
Claude breaks this cycle forever. Maintenance becomes just a command. Reorganizing the entire Vault is one prompt. Migrating from Notion? Processing export files, adding properties, and restructuring into a new system—all automated.
At the end, the original article mentions Vannevar Bush's Memex (1945). A personally curated knowledge store where connections between documents are as valuable as the documents themselves—Bush envisioned this, but what he couldn't solve was "who does the maintenance."
Now, the answer is here.
■ Summary
・The LLM Knowledge Base proposed by Karpathy is an approach to "giving AI long-term memory"
・Level 1 can be started with just copy-pasting between Obsidian + Claude chat
・Level 2 is a 3-layer architecture (raw / wiki / CLAUDE.md) and 4 cycles (Ingest / Compile / Query / Lint)
・Level 3 is 5 stages of automation (CLI → Slash commands → Schedule → GitHub Actions → Agent Skills)
・What should go in the Vault is "everything you consumed and lost in the past year"
・AI completely takes over "maintenance," the biggest reason past second-brain tools failed
・You can start immediately with community-made plugins (wiki-skills, etc.)
Source: @hooeem
https://x.com/hooeem/status/2041196025906418094
━━━━━━━━━━
To those who found this article even slightly helpful.
Claude Code Studio @ Japan (@ClaudeCode_love) is
an account run by three Claude Code enthusiasts.
We post daily about practical CLI utilization and automation.
We are currently co-developing AI agents with listed companies.
Our usual content 👇
・Real product development cases using Claude Code and Claude
・Organizing Claude Code usage / Vibe Coding / development trends
・Latest information on Claude Code from overseas
From development philosophy to design, implementation, and improvement,
we summarize things with the premise of "putting a working product out into the world," not just "finishing the build."
If you're interested, please follow and check us out 👀
