Instead of Obsidian, Use Claude Code

An extension of my memory, an external brain where AI remembers and connects on my behalf

"After reading all kinds of materials stored in my iCloud, can you judge and tell me who my closest friend is?"

Last week, I asked AI this question. It was after I had it read ten years of diaries, individual notepad notes, Excel files, and contacts stored in iCloud all at once. About three minutes passed. Claude Code gave its answer.

"It is OO."

It was not a wrong answer. Other people also appear often in the diary, but that is in the context of work meetings. OO was different. A phone call on a Sunday with nothing to do, the first call made during a busy period, a night spent together all the way to karaoke, the worry written in the diary when my father collapsed. AI read that grain.

I realized something. The time I had spent in Obsidian making links, adding tags, and staring at graph view suddenly felt empty. What I wanted was not a pretty knowledge graph. It was another brain that remembers my memories for me.

Obsidian Is Still Manual Work in the End

Obsidian is an excellent tool. Because it is Markdown-based, it is light; because it stores locally, there is no privacy concern; and the plugin ecosystem is solid. But there is one decisive problem.

Every connection must be made by hand.

Only when you link [[name]] does it connect to that note. Only when you add the #friend tag does it appear in the friend list. If you wrote "called OO" in a diary entry on some day in 2021, Obsidian does not know why that sentence matters or what situation you were in that day. That is an area where you have to manually annotate.

There are ten years of diaries. There are thousands of notes. There are hundreds of contacts. How many months would it take to link all of this, tag it, and organize the relationships? And even if you invested those months, could Obsidian answer the question, "Who is my closest friend?"

No. Obsidian is a search tool, not a reasoning engine.

Claude Code + Storage Indexing, This Is DualBrain

The method I chose has four steps. In fact, everything below can be done by instructing Claude Cowork or Claude Code.

First, organize files into a form Claude Code can read. Convert Hangul (HWP) files, PDFs, Word documents, and KakaoTalk backup files into Markdown (.md). CSV and Excel (XLSX) files can be read as they are, so leave them alone. Claude Code writes the conversion script. A script that ties together Python libraries such as hwp5txt, pdfplumber, and python-docx and converts an entire folder can be completed in 30 minutes.

Second, gather the converted files in a specific iCloud folder. In my case, under ~/iCloud/ai_workspace/memory/, I divided diaries, notes, contacts, project memos, and other materials into subfolders. You do not have to gather everything in one place. Even if files are scattered across iCloud, there is a way.

Claude Code also does the indexing work here. The only thing a person needs to do is give one natural-language instruction. Whether there are 10 folders or 50, you do not need to issue separate commands for each folder.

Tell Claude Code, "Go through all subfolders under ~/iCloud/ and create an index for each folder. For Markdown files, extract the filename, date, and key keywords. For CSV or Excel files, organize the sheet names, column structure, and date ranges. When everything is done, combine them into one master index called master_index.md."

Claude Code identifies the folder list, visits each one, opens and reads files, summarizes them, creates folder-by-folder index files, and finally creates the master index. A person does not need to open Excel or manually organize a list.

A master index made this way becomes powerful later. When a question comes in, Claude Code does not open thousands of files one by one. It scans the master index first and decides which folder to read deeply. Time and token cost drop sharply.

Third, connect a filesystem server to Claude Code. This is the core. Once Claude Code can directly access the local file system, it can read and analyze thousands of files in real time. You can give it all of iCloud as the working folder. Or you can use a collection folder you choose.

Fourth, ask questions. In natural language. "Who did I meet most often in 2023?" "On what date did I first contact the client for that project?" "When was Mom's birthday again?" What is someone's phone number? I met them, but why did I meet them? Who came to my son's wedding? Claude Code digs through the files and finds the answer. You can ask it to find a report you wrote in the past on a specific topic, or ask it to extract only that report's table of contents.

This is the Dual Brain. An extension of my memory. An external brain that remembers what I forgot and connects what I failed to connect.

What Can the Dual Brain Do?

These are use cases I actually tested.

Relationship analysis

"Who do you think is my closest friend?" In response to this question, the AI did not count frequency. Even if someone appeared 50 times in my diary, if all 50 mentions were in the context of work meetings, it did not classify that person as a 'friend.' If another person appeared 30 times, but the mentions included 'private context' such as Sunday calls, karaoke, or worries about a family illness, it rated the intimacy higher. How would this be implemented with tags in Obsidian? Even if you add tags such as #friend #work #family, it is hard to express "a relationship that began through work but later became friendship" with tags. AI reads context.

Restoring memory along the timeline

"What happened in March 2021?" March of that year was a busy month for me. There were three business trips and a new project kickoff. In the middle of that, my diary records calls to three acquaintances. AI found this and even added the interpretation, "key relationships you contacted first even during a busy period." Why I called those people was context I had forgotten until I asked two years later.

Tracking appointments and history

"When did I last meet the client CEO?" Business meeting history. It may be in the calendar, but calendars often only say "meeting" without the other person's name. In the diary, it says, "Lunch with CEO OO, and a new project came up there." AI cross-checks the calendar and diary to find the exact date and context.

Pattern recognition

"At what time of day did I write the most?" "What is the difference between the months when I exercised consistently and the months when I did not?" "What expressions do I often use when I am stressed?" This is meta-analysis of oneself. What would take three days if I read 10 years of diaries myself, AI scans within 5 minutes and extracts patterns. When it analyzed, "When you are stressed, your sentences get shorter, and commas increase instead of periods," I got chills. It was the moment AI told me about a version of myself I did not know.

Recovering lost information

"What was the name of that person I met three years ago? They went to a top university and said they were doing an AI startup." Only fragments of memory remain. I cannot remember the name. With Obsidian? I cannot search because I do not know the name. With AI? It scans diaries with fragmentary keywords and narrows down candidates. The answer comes back: "There is a record about that person in the diary entry for July 15, 2021."

Discovering information inside files

Finding a specific phone number in an Excel file, organizing contract terms scattered across hundreds of documents at once, or selecting only files containing a certain keyword from an old project folder. AI shines in these chores.

Go deeper, and it can do things like this too

Finding writing material and fact-checking

When writing a column or book manuscript, there comes a moment when you need to find, "What did I write about this topic before?" The Dual Brain scans all past manuscripts, extracts relevant sentences, and cross-checks whether numbers or sources you cited before conflict with what you are writing now. If it tells you, "In a 2023 column, you wrote this figure as 30%, but in a 2024 memo, you wrote it as 25%," you can catch factual errors in advance.

Support for legal work

For lawyers, the Dual Brain becomes a practical work tool. "Find advisory records related to a certain company that mention personal information issues." "Extract the list of precedents cited in last year's case brief." For a lawyer with hundreds of advisory memos and legal opinions piled up, AI catches contextual links that keyword search cannot catch.

Tracking meeting minutes and promise fulfillment

"Extract every item from the last three months of meeting minutes that we said we would do next." Things said in meetings such as "Let's discuss that next time" or "I will research and share it." Most are forgotten. The Dual Brain reads all meeting minutes and creates a list of unfulfilled promises, with dates, owners, and original context included.

Health record analysis

Diary entries may have scattered records such as "headache," "couldn't sleep," or "back hurts." If you ask the Dual Brain, "Organize only this year's health-related records in chronological order," it creates a symptom timeline to show a doctor before visiting the hospital. It even catches correlations such as "Headaches were concentrated in March and July, and both periods had high overtime frequency."

Children's growth records and education history

For parents who have recorded a child's growth, the Dual Brain is a gift. "When was the first step?" "Organize vaccination records." "Show academy changes together with grade changes." It combines parenting diaries, hospital records, and text from school notice photos to reconstruct the child's growth timeline.

Cross-analysis of financial records

"Were there any overseas business trip expenses last year that I was not reimbursed for?" Receipt photos, expense reports, and card statements are scattered across different folders. The Dual Brain cross-checks them and finds missing items. Before tax filing, how convenient would it be if it told you, "This amount was processed as an expense, but there is no receipt file"?

Reading records and tracing intellectual lineage

"Collect everything I have noted about 'AI ethics,' including which book I read it in and when." Even if highlights from books, reading notes, and blog reviews are scattered, AI ties them into one intellectual lineage. When writing a new piece, rediscovering thoughts that my past self had already organized is quite thrilling.

Reconstructing travel records

"What was the name of the restaurant I went to in Japan a few years ago? That ramen place in an Osaka alley." Combine the dates of photo files, card payment records, and fragmentary diary mentions, and the answer appears. Something like, "There is a card payment record in Osaka on that date, and on the same day your diary says, 'Ramen I waited in line for, the broth was hot.'"

How to Automate File Conversion

The first hurdle in building a Dual Brain is file format. A large share of files piled up in iCloud are HWP, PDF, and DOCX. You need to convert all of them to Markdown so Claude Code can read them. In fact, you can ask Claude Cowork or Claude Code to do this too.

For converting Hangul (HWP) files, use pyhwp or hwp5txt. On Mac, another method is to convert HWP to ODT with LibreOffice in headless mode and then extract text. It is not perfect, but most body text survives.

For PDFs, pdfplumber is a safe choice. PDFs with a text layer are extracted almost perfectly. Scanned PDFs require OCR, and adding the Korean language pack to Tesseract OCR gives about 80% accuracy. It is not perfect, but it is enough for search.

DOCX converts easily with python-docx. Tables and images disappear, but the body text transfers to Markdown as is.

CSV and Excel (XLS/XLSX) require a different approach. There is no need to convert them to Markdown. CSV is already a text file, so Claude Code reads it directly. XLSX is binary, but if you open it with Python's openpyxl or pandas, you can pull out the contents immediately. Core data such as contacts, expense history, and project management sheets are often in Excel, so leaving this out of the index creates a hole in the Dual Brain. If you include "sheet name, column structure, row count, date range" when creating the index, later, for a question like "Get the total taxi expenses for 2023," it can immediately find which file to open.

Ask Claude Code, "Write a Python script that converts every HWP, PDF, and DOCX file in this folder to Markdown, and creates a sheet-by-sheet structure summary for CSV/XLSX files." Within 30 minutes, you will have code that runs. Once written, run it whenever new files come in.

The Role of the Index, and Real-Time Analysis

Obsidian creates an "index" with the Dataview plugin. If you add YAML frontmatter at the top of files and write metadata such as date, tags, and category, Dataview queries it and shows it as a table.

The Claude Code method also works better if you create a summary index in advance. That is exactly why we converted files to Markdown earlier and created folder-level indexes and a master index.

With that index, Claude Code can go straight to the necessary folder without reading thousands of files every time. It has the same effect as Obsidian building a structured index in advance with YAML frontmatter and Dataview and pulling it out instantly. If you have Claude Code write an automation script that says, "Every December 31, create a list of the top 20 people I met this year and save it to index.md," and put the result into the Obsidian vault too, you can get both instant answers and real-time analysis.

On the other hand, for questions that cannot be designed into an index in advance, such as "Is there anyone I grew distant from between 2020 and 2023?", or for analyses that do not exist until the question comes to mind, Claude Code is overwhelmingly strong. The two are not competitors. They have different roles.

If you do not create an index, it has to read thousands of files from beginning to end every time, and then token cost becomes unmanageable. In practice, Claude Code first scans file names and paths, then selects only the files that seem related to the question and reads their contents. If needed, it also creates summaries and uses them as an intermediate step. From the user's perspective, "I ask and the answer comes out," but behind the scenes, fairly sophisticated selection and summarization are running. That is why pre-indexing matters.

Hardware and Models Determine the Dual Brain's Ability

The Dual Brain's ability is determined by two things.

Hardware performance. To run Claude Code locally, you need enough RAM and a fast SSD. In my case, I use a MacBook Pro M5 Pro as my main machine. I also have an M4 Mac Mini, and I run Ubuntu on an Intel Mac Mini as a server. If the number of files is small, an M4 is enough, but to analyze thousands of files at the same time, you need plenty of memory. SSD speed also matters. Every time Claude Code reads a file, disk I/O occurs. With an HDD, it is too sluggish to use.

The underlying language model. Claude Code uses Claude Sonnet by default. If you run OpenClaw locally, performance varies drastically depending on which model you attach. Gemma 4 on a local Mac still has shallow reasoning depth. If you attach Claude Opus, it is slow but answers complex questions accurately.

To answer the question "Who is my closest friend?" properly, context understanding is required. A small model cannot do it. If the model is small, it gives a foolish answer such as "Based on name occurrence count, OO is your closest friend."

Put it this way.

• Small model + slow hardware: keyword matching level. It is embarrassing to call it a Dual Brain.

• Small model + fast hardware: fast search. Reasoning is still weak.

• Large model + slow hardware: deep reasoning is possible, but answers take a long time.

• Large model + fast hardware: this is the real Dual Brain. You ask a question, take a sip of coffee, and the answer appears.

What If We Use Obsidian and Claude Code Together?

Obsidian is still a good tool. For writing, organizing ideas, and managing projects, few tools are as good as Obsidian. As a Markdown editor and as a local note app, Obsidian still has its place.

What I am talking about is Obsidian's limit as a "second brain." Obsidian connects only what I connect. It recognizes as tags only what I tag. I cannot throw 10 years of diaries at it and say, "Analyze this on your own."

Claude Code plus file indexing can do that. AI makes the connections. AI finds the patterns. I only ask questions.

What if the two are used together? Write in Obsidian. Analyze and search with Claude Code. If you connect the Obsidian vault folder to Claude Code's MCP filesystem, Claude Code analyzes notes written in Obsidian in real time. It is a good combination. But as Claude Code becomes able to analyze and synthesize all files perfectly as long as indexing is done well, even if files are left scattered without order, Obsidian's value will likely decline.

Why Opus 4.6 and High-End Hardware?

The quality of the Dual Brain is ultimately decided on two axes: how intelligent a model you use, and how quickly that model processes files.

A top-tier intelligence model is good. Think again about the question, "Who is my closest friend?" A small model counts name appearances. The person who appears 50 times is ranked first. This is something even the grep command can do.

Opus 4.6 approaches it differently. It reads the entire diary and understands context. "This person appears only in work meetings." "That person contacts you even on weekends, talks about family, and calls first when things are hard." It makes these distinctions by itself. It reads depth of relationship, not frequency.

Once I asked this question: "Is there anyone I grew distant from between 2020 and 2023?" Haiku only tracked frequency change. "OO appears 12 times in 2020 and 3 times in 2023." That is true, but superficial. Opus 4.6 was different. "The tone of conversations with OO changed. In 2020, there were many expressions such as 'dinner together' and 'let's meet on the weekend,' but after 2022, the expressions change to 'I should contact them' and 'I haven't seen them in a while.' The qualitative change stands out more than frequency." I got chills. It was the moment AI pointed out a change in a relationship that I had not consciously noticed.

The more complex the reasoning required by a question, the more model size matters. "How has my writing style changed from three years ago to now?" "What patterns appear when I am stressed?" "Have I ever changed my position on a certain topic?" To answer questions like these properly, the model must read thousands of files, sort them chronologically, detect subtle changes in tone, and extract patterns. Local models at the 7B or 13B level are not enough.

Hardware is the same. The unified memory of the M5 Pro has a reason. When Claude Code reads files, it must load the full contents into memory. With large memory, even 10,000 files can be handled. SSD speed is also felt directly. A task that takes 3 seconds on an NVMe SSD takes 30 seconds on an external HDD.

In the end, the Dual Brain formula is this.

Dual Brain quality = model intelligence x hardware speed x data volume

If any one of these three variables is lacking, the experience is cut in half. Use a small model, and reasoning is shallow. Use slow hardware, and answers do not come. Have no data, and there is nothing to analyze.

I use a combination of the Opus 4.6 API and a MacBook Pro M5 Pro. It costs money. The monthly cost is not small, and the MacBook price is not light either. That is also why I use the Claude Max plan. After experiencing a 3-minute search across 10 years of memory and seeing AI restore the context of relationships I had forgotten, that cost does not feel wasted.

How to Start

No need to make it complicated.

• Create a folder in iCloud, such as ~/iCloud/ai_workspace/memory/.

• Gather diaries, notes, and memo files under that folder. You can divide them into subfolders. Or not. You can simply index and search all of iCloud or your storage device as one whole. It just consumes more tokens and returns results a little later.

• If there are HWP, PDF, or DOCX files, convert them to Markdown. Ask Claude Code to write the script.

• Connect an MCP filesystem server to Claude Code. In the terminal, one line is enough: claude mcp add filesystem -- npx @modelcontextprotocol/server-filesystem ~/iCloud/ai_workspace/memory. To share it with team members, create a .mcp.json file in the project root.

• Ask questions. "Who did I mention most often in 2022?" "Find only the memos related to a certain project that mention the client."

At first it feels awkward. You wonder, "Can I even ask this?" But once you use it, you know. It is more accurate than my memory, faster than my search, and deeper than my analysis.

When AI reads a diary or memo file written 10 years ago and says, "At that time, you were worrying about this. This is what you were thinking." In that moment, you feel what the Dual Brain is. And the depth of that feeling is determined by the model and hardware you choose.

There is, however, one premise behind all of this: records. You need 10 years of data files to search 10 years of memory. You need to leave memos before patterns can be extracted from them. No matter how smart AI becomes, it cannot replace the brain of someone who has recorded nothing. The most primal fuel that determines the performance of the Dual Brain is neither model nor hardware. It is the fragments of life that a human wrote down by hand.