“Am I really using AI to its full potential?”
“Why is there such a performance gap between the AI others use and my own?”
“I’m tired of entering the same background info every time I use AI...”
These are common frustrations. Today, we’re going to solve them.
The reason AI doesn't work exactly as you want isn't a matter of the AI's intelligence; it's because you haven't designed its "memory."
Have you experienced this while using AI like Claude Code?
- Explaining your job or interests to the AI every single time.
- Bookmarks piling up, but you can't find an article you saved a month ago.
- Someone else getting better results from the same Claude model.
- Output quality remaining inconsistent despite your best prompt engineering.
By reading this article to the end, you will fully understand how to build a "Second Brain" using Claude Code × Obsidian, from the philosophical background to specific setup steps, and be ready to start from zero today.
The combination of Claude Code × Obsidian has become incredibly hot recently. While it has been used for a while, the momentum overseas is particularly massive right now.
For instance, triggered by a proposal from former OpenAI founding member Andrej Karpathy, Obsidian CEO Steph Ango himself has released AI integration skills. Key industry figures are moving one after another.
Related overseas articles are hitting millions of views. Total views exceed 12 million, with over 80,000 bookmarks.
My previous article summarizing these comparisons received bookmarks from over 500 people when I only had 10 followers, proving how useful this is.
UT ClaudeCode Research Institute
@ClaudeCode_UT
·
Article
[Complete Archive] Thorough Explanation on How to Link ClaudeCode with Your Second Brain
“The AI agent isn't moving as I expected... !💢”
“AI is totally useless... !😎”
This isn't the AI's fault; it's how you use it...
3 47 425 150K
Before we start, I have two requests:
- Save this article and set aside 20 minutes this week to read it.
- Share it with anyone you know who is interested in AI.
Furthermore, I want you to actually try it. Even just telling Claude Code to "read this article and execute it" is fine.
There is a 100x difference in the AI era between those who save and forget and those who take this extra step.
I will break down these overseas insights along with tips I've discovered through my own use so that even a complete beginner can start from scratch!
Key source articles are here:
▶︎ @defileo: Claude + Obsidian have to be illegal
▶︎ @NickSpisak_: How to Build Your Second Brain
▶︎ @sourfraser: Claude + Obsidian = A true AI employee
Obsidian × Claude Code is No Longer Just for "Geeks"
In the overseas AI scene, one combination is attracting abnormal attention: Obsidian × Claude Code.
Articles combining these two are recording millions of views. Six major articles alone totaled 12.4 million views and over 80,000 bookmarks.
Moreover, this isn't just hype from general AI influencers. Obsidian CEO Steph Ango developed AI agent integration skills himself and released them on GitHub, gaining over 25,000 stars. The creators themselves are evolving products with the premise of "using them with AI."
One survey shows 44% of experts cite AI as the most important technology for knowledge management. Lex Fridman also uses Obsidian.
One overseas user said: Every morning when I open my PC, the AI starts working from where it left off yesterday without any explanation. I don't have to re-explain my work context every time.
This is becoming the new standard for intellectual labor, not just an experiment for geeks.
Most People Use AI as an "Amnesiac Temp Worker"
Overseas AI consultant sourfraser put it perfectly: "Most people use AI like an amnesiac temp worker."
Imagine how inefficient it would be if you had to explain who you are, what your job is, and what you want every single morning you go to work.
As a member of this research institute, I was in that exact state... Every time I used Claude Code, I would type: "I am researching business applications of AI agents, and recently I've been interested in..." Now that I've designed its memory, this explanation is unnecessary. The AI naturally proceeds with work while maintaining context.
sourfraser shared this experience: "I completely forgot what we decided on a call. Two days later, I asked Claude Code, and it retrieved it accurately."
Even using the same Claude model, the output changes completely based on the quality of the "memory" you provide.
AI without memory is just a "search engine with a personality." AI with memory becomes a colleague who truly understands your work.
The Era of Competition Based on "Memory Design," Not AI Performance
The focus of AI utilization has shifted significantly over the years:
- 2023-2024: Which AI tool to use
- 2024-2025: What prompts to write
- 2025-2026: What context or harnesses to utilize
- 2026 onwards: How to design and operate memory
Most people are still at the "optimizing prompts" stage. But prompt optimization is a one-time effect. You have to write a good prompt every time.
In contrast, "memory design" has a compounding effect. The more memory you give the AI, the more its output accuracy increases automatically.
Knowledge grows with "compound interest."
There will be an unbridgeable gap in accumulated memory between someone who starts today and someone who starts in six months. The era of competing on AI "performance" is over. Since we use the same models, the difference lies in "what you provide as memory."
Knowledge Management Has Always Fought the "Sustainability Problem"
Attempts to build a "Second Brain" have been repeated for decades. And almost all have failed for the same reason. Let's look at some methods:
[Folder Organization]
The method everyone tries first. Saving files by topic. You might remember filenames like "Project_Material_v2_Final_Fixed.pdf." It's a management hell...
Once you exceed 100 files, "where was that file?" becomes a daily occurrence.
[Bookmarks]
Browser bookmarks. You can save articles with one click. But look back: how many of those saved articles did you actually re-read? Searching for the same article with the same keyword a month later is the trap of "satisfaction through saving."
[PARA Method]
Tiago Forte's classification: Projects / Areas / Resources / Archives. Good for tasks, but weak for pure knowledge accumulation. It organizes "what to do" but doesn't accumulate "what you know."
[Zettelkasten]
Devised by German sociologist Niklas Luhmann, leading to over 70 books. It's brilliant, but the manual maintenance load is too heavy for normal people to sustain.
[Evergreen Notes]
Proposed by Andy Matuschak. It's quite cool but has the weakness of easily falling into perfectionism.
defileo describes the pattern where a "Second Brain" dies:
"Organize and start. Maintenance piles up. Skip it. Quality degrades. Return to messy notes. Try again 6 months later. Repeat."
The common weakness of all traditional methods was the "premise of human maintenance."
In 1945, Vannevar Bush envisioned "Memex," a personal knowledge device. A device to solve the information maintenance problem. But even Bush couldn't solve the problem of who does the maintenance.
Then came the "LLM Wiki" proposed by Andrej Karpathy, former OpenAI founder and Tesla AI head.
AI, including Claude Code, reads sources, automatically builds a Wiki, and takes over maintenance.
This is the first practical answer to an 81-year-old problem.
Actually, the LLM Wiki incorporates all the best parts of historical PKM methods:
- Zettelkasten's "one page per concept + interconnection via links + index"
- Evergreen Notes' "concepts evolve as they are used"
- MOC (Maps of Content) "creating a global map of knowledge with an index"
The only difference: Is it done by a human or an AI?
AI automatically creates the Zettelkasten structure, grows concepts like Evergreen Notes, and keeps the index updated like MOC. AI has fundamentally solved the problem where past methods ended with "it's good but unsustainable."
Why Obsidian is Being Chosen
You might think, "Isn't Notion fine for knowledge management?"
Notion has its strengths. But Obsidian is chosen for the combination with Claude Code for three reasons:
[Plain Text: AI Can Read and Write Directly]
Obsidian files are all Markdown. Claude Code is an agent that can directly manipulate the file system, so it can read, write, and add links to Obsidian notes as they are. No need for APIs or plugins.
Obsidian is simply a system of md files connected by links. This simple, flat structure has the best compatibility with AI.
[Local: It's Fast]
Obsidian data is all on your PC. No waiting for cloud sync. Data won't disappear if a service ends. The speed at which Claude Code reads and writes local files is the fastest.
[Tool Evolving with AI in Mind]
Just recently, Obsidian CEO Steph Ango developed AI agent integration skills and released them on GitHub (25,000+ stars). The creators are evolving the product with the premise of AI usage.
As a side note, Obsidian's strength lies in simplicity, so be careful not to make adding plugins the goal itself. That's the trap of self-purpose.
Personally, I'm all for Obsidian, but considering operational lightness, other approaches are also effective. I'll introduce them below.
There is More Than One Way to Give AI Memory
There are currently three major approaches to giving AI memory. They are independent and used based on the purpose.
① LLM Wiki Method (Obsidian × Claude Code)
The method proposed by Karpathy and explained in this article. Put raw materials in an Obsidian vault, and Claude Code automatically structures them as a Wiki. The Wiki grows as you use it, increasing accuracy. This has the strongest knowledge compounding effect.
However, token consumption can be high. As sources accumulate, the cost per question rises, so it's important to use it with a sense of cost. Suitable for long-term tracking of a theme, researchers, or specialists.
② NotebookLM Method (Google)
A completely different approach. You can ask questions immediately after uploading sources to Google's NotebookLM. It's incredibly easy.
However, knowledge does not accumulate. It's a disposable type for each project. Perfect for "I want to ask about this material right now," but it doesn't create long-term knowledge compounding. It's not a competitor to Obsidian × Claude Code; the use case is different.
③ Skills / CLAUDE.md Method (Claude Code Alone)
A method of writing business context in Claude Code's CLAUDE.md. It stays within the project folder without using Obsidian. You can define "how you want the AI to act." It's the closest method to "using" rather than just "storing."
There is no single correct answer. It's realistic to use them based on purpose. Many people use them together.
This article focuses on the LLM Wiki method, which is the most talked about and has high compounding effects, explaining the steps to build a "Second Brain" with Obsidian × Claude Code.
From here is the practical part. I'll explain Claude Code, Obsidian, and how to combine them.
Let's Start Claude Code
If you're already using Claude Code, you can skip this. Claude Code is an agent-type AI tool provided by Anthropic.
How is it different from ChatGPT or Claude chat? The biggest difference is that it can directly read and write local files.
Normal AI chats reset when the conversation ends. Next time, you have to explain from scratch. Claude Code is different. It can read files on your PC, create new ones, and rewrite existing ones. In other words, it can hold "memory" as files.
Just by placing a file named CLAUDE.md in your project folder, the AI automatically reads it. If you write "who you are," "what you value," and "how you want it to behave," you won't need to re-explain every time.
Pricing Plans
There are several plans. There's info saying "Claude Code can't be used with the Pro plan," but since this changes, as of now, I can use it with the Pro plan.
- Pro ($20/month): Cheapest plan. But you'll hit usage limits quickly.
- Max 5x ($100/month): 5x usage. Recommended for serious use.
- Max 20x ($200/month): 20x. For heavy users.
Starting with Pro and upgrading to Max 5x when needed is realistic.
Installation
Claude Code can be used via CLI (terminal) or a desktop app. If you're not used to the terminal, start with the desktop version. CLI installation requires Node.js v18+. Download the "LTS version" from nodejs.org.
1npm install -g @anthropic-ai/claude-code
Once installed, just type claude in any folder to start. You can also ask the Claude app for instructions on how to proceed.
Obsidian is Free and Ready to Use Today
Simply put, Obsidian is a free note-taking app.
Its features are simple: saved locally as Markdown, no subscription, no lock-in. Data is entirely in your hands. By connecting notes with "[[links]]", a knowledge network forms naturally. The graph view allows you to visualize the whole picture.
Installation and Initial Setup
- Download and install from the Obsidian official site.
- Select "Create new vault." A Vault is a folder to store notes.
- Specify the save location (Desktop, Documents, etc.).
First File: Memory File
Once the Vault is created, make one file: "Memory.md". Write:
- What is your job?
- What projects are you working on?
- What tools do you use often?
- What are your work goals?
- What criteria do you value?
sourfraser calls this an "onboarding document for a new employee." You write a document explaining yourself for the AI. This is the starting point.
20 minutes is enough. It doesn't have to be perfect. You can even have Claude Code set this up. I recommend doing it yourself once to understand what's happening.
Another useful file is "Home.md," acting as a hub for the Vault with links to major notes.
Vault Folder Structure
Use the 3-folder structure recommended by Karpathy:
``bash
vault/
├── .raw/ ← Raw materials. Articles, PDFs, notes.
├── wiki/ ← Knowledge pages generated/maintained by AI.
└── outputs/ ← Deliverables. Reports, drafts, etc.
``
Throw everything into .raw/ without organizing. AI will handle the organization.
Setting the Vault location to Google Drive allows automatic sync across machines.
obsidian-skills and claude-obsidian Make Everything Work
For those who find Claude Code and Obsidian confusing and want to try it quickly, here is the perfect info. Two open-source projects exist to combine them.
■ obsidian-skills: The Foundation Layer
- Author: Obsidian CEO Steph Ango
- Stars: 25,720 (as of 4/22)
- License: MIT
- GitHub: kepano/obsidian-skills
A skill set that teaches AI "how to use Obsidian." It includes 5 skills:
- obsidian-markdown: Teaches correct Markdown writing.
- obsidian-bases: Database view operations.
- json-canvas: Visual canvas operations.
- obsidian-cli: Vault read/write/search commands.
- defuddle: Extracts clean Markdown from web pages (saves tokens by cutting ads).
■ claude-obsidian: The Application Layer
- Stars: 2,570 (as of 4/22)
- License: MIT
- Based on Karpathy's LLM Wiki pattern.
- GitHub: AgriciDaniel/claude-obsidian
A skill set for having AI "build and maintain a second brain." Includes 10 skills. Major commands:
- /wiki: Initial setup. Automatically builds Vault structure.
- ingest [file]: Reads material and generates 8-15 Wiki pages.
- /save: Saves the current conversation as a Wiki note.
- /autoresearch [topic]: Runs 3-5 rounds of web research on a topic.
- /canvas: Creates/manipulates visual boards.
- lint the wiki: Health check for broken links or contradictions.
There are 6 Wiki Modes: Website, GitHub, Business, Personal, Research, Book/Course.
Installation
Method 1: Install as a plugin
``bash
claude plugin marketplace add AgriciDaniel/claude-obsidian
claude plugin install claude-obsidian@claude-obsidian-marketplace
``
Method 2: Clone the repository
``bash
git clone https://github.com/AgriciDaniel/claude-obsidian
cd claude-obsidian
bash bin/setup-vault.sh
``
Running setup-vault.sh sets up graph view colors, CSS, and folder structures. Open this folder as a Vault in Obsidian to start.
The Resulting Vault Structure
Running /wiki creates:
``bash
claude-obsidian/
├── .raw/ ← Raw materials (Claude reads only)
├── wiki/
│ ├── index.md ← Table of contents (auto-updated)
│ ├── hot.md ← Recent context cache (~500 words)
│ ├── log.md ← Operation log
│ ├── overview.md ← Vault summary
│ └── {topic}/ ← Wiki pages (1 concept per page)
├── _templates/ ← Note templates
└── _attachments/ ← Images/PDFs
``
This structure solves the sustainability problem: AI automatically creates and maintains everything. You just put materials in .raw/.
Recommended plugins:
- Templater: Auto-apply templates.
- Obsidian Git: Auto-commit every 15 mins for version control.
- Web Clipper: Browser extension to save pages to
.raw/with one click.
Daily Operation: "Input → Ask → Grow"
① Input
Drop materials into .raw/ and type ingest [file]. AI generates 8-15 Wiki pages with an average of 12 wikilinks each, making the network dense naturally.
② Ask
Ask the Wiki: "What do you know about [topic]?" Claude Code reads hot.md and index.md to find relevant pages. This keeps token costs constant even as the vault grows to thousands of pages.
③ Grow
Use /save to store conversation insights. defileo says: "Humans choose sources, ask good questions, and think about meaning. Claude does the rest."
KanikaBK even automates morning briefings where Claude Code scans Inbox and DailyNotes to extract tasks and update idea lists. Use lint the wiki every 10-15 ingests to fix broken links or contradictions.
Experience the "Compound Interest"
sourfraser describes the growth:
- Week 1: Basic responses.
- Month 1: Grasps client info, team structure, and processes.
- Month 2: AI starts pointing out things humans missed.
Most People Stop at "Satisfaction from Building"
Many people spend time on setup and then never use it. Avoid these patterns:
- Spending too much time on plugins.
- Trying to design a perfect folder structure.
- Obsessing over the quality of the first notes.
defileo says: "The vault doesn't have to be perfect. It has to be real." Also, watch out for "compounding errors"—if you save wrong info, the AI's answers will warp. Use lint the wiki to maintain quality.
Realize the Gains in the First Month
Setup is fast—it took me less than 10 minutes. Don't try to be general at first; focus on one theme (e.g., sales prep, webinar materials, YouTuber links) to experience the effect quickly.
1-Month Roadmap:
- Week 1: Setup Claude Code, Obsidian, and skills. Write Memory.md. Input 3 materials.
- Week 2-3: Habituate inputting articles and notes. Ask the Wiki questions.
- Week 4: Experience AI instantly answering "that thing I researched last week."
In 1945, Vannevar Bush dreamed of Memex. 81 years later, it's becoming reality via Claude Code × Obsidian. Designing memory is a compounding structure. Those who start early will create an uncatchable lead.
Summary
- The next phase of AI is "Memory Design." Prompts are one-time; memory compounds.
- AI has finally solved the "sustainability problem" of knowledge management.
- Obsidian's plain text is perfect for AI. Claude Code integration is seamless.
obsidian-skillsis the foundation;claude-obsidianis the application.- Daily cycle: Input → Ask → Grow. Costs stay constant as the vault grows.
- Don't just build it; use it. The first notes can be low quality.
- Start this weekend, and in a month, the AI will be your most knowledgeable colleague.
If this was helpful, follow @ClaudeCode_UT (UT Claude Code Research Institute). We provide practical techniques and unique overseas insights daily to make your Claude Code stronger!
