A second brain for Claude, without the forty-step setup
What Karpathy's LLM wiki really is, what Claude already remembers on its own, and the smallest second brain that works: three files and one paragraph.
PS
Prashant Shaw
Founder, Sumibako
8 min read
Search for "second brain for Claude" and you will find setups with forty steps. Nine plugins, a vector database, a cron job called the heartbeat, and a Slack bot that drafts your replies in your voice. They are genuinely impressive. They are also how most second brains end: as one long weekend, followed by a folder called brain-v3-final that nobody opens again, least of all its owner.
This post is the shorter road. What the idea actually is, from the document that started the current wave of it, which turns out to be calmer than most of the setups it inspired. What Claude already remembers about you without being asked. The smallest version that works, which is three files and one paragraph. And the catch that every folder-based setup hits the first time you pick up your phone. Facts are linked to their sources and were checked on 3 October 2026. The last section is about the thing I build, and it says so.
🧠 The idea, from the source
On 4 April 2026 Andrej Karpathy posted a gist called LLM Wiki. It is about two thousand words and describes a pattern rather than a product. The core of it: instead of a model searching your raw documents every time you ask something, it "incrementally builds and maintains a persistent wiki", a set of Markdown pages that it writes, cross-links and keeps current, so that knowledge is "compiled once and then kept current, not re-derived on every query".
It has three layers:
Raw sources. Articles, papers, notes. The model reads them and never changes them.
The wiki. Pages the model writes and owns: summaries, topics, comparisons.
The schema. One document, his example being a CLAUDE.md for Claude Code or an AGENTS.md for Codex, that tells the model how the wiki is organised and what to do when you hand it something.
And three things you do with it: ingest a new source, query the wiki, and lint it now and then for contradictions, stale claims and orphan pages. His one-line summary is the best sentence in the whole genre: "Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase."
The reason it works is in there too: "Humans abandon wikis because the maintenance burden grows faster than the value." I felt that one personally. I have abandoned four, one of them about how to keep a wiki.
Two details in the gist get lost in the retellings. He says an index file "works surprisingly well at moderate scale (~100 sources, ~hundreds of pages) and avoids the need for embedding-based RAG infrastructure". So you do not need a vector database to start, and possibly ever. And he calls the document "intentionally abstract", with everything in it optional. The forty-step versions are fan fiction. Some of it is excellent fan fiction. It is still fan fiction.
🐟 What Claude already remembers
Before building anything, it helps to know what Claude keeps on its own, because it is more than people expect and less than they assume.
In the Claude apps, on the web, desktop and phone, memory is on by default for the Free, Pro and Max plans, and "each project has its own separate memory space". In Claude Code there are two kinds: CLAUDE.md files, which you write, and auto memory, which Claude writes for itself. Auto memory is a folder with a MEMORY.md index and one file per note, in four kinds: who you are, corrections you gave it, the state of a project, and where to find things. The first 200 lines of the index, or 25KB, are loaded at the start of every session.
Then the sentence that matters: "Auto memory is machine-local." The docs go on to say the files "are not shared across machines or cloud environments".
So Claude remembers you the way a family of goldfish would, each in a separate bowl: one per chat project, one per repository, on each laptop. Every bowl is useful. None of them is a second brain, for one reason. Those memories are Claude's notes about you, written for Claude. A second brain is your notes, which Claude happens to maintain, and which you can read, search and hand to someone else.
🧰 The smallest second brain that works
Three files and one paragraph. Make a folder, say ~/brain, with:
1️⃣ index.md: one line per page, with a one-line summary. The model reads this first, every time.
2️⃣ log.md: append-only, one dated line per change. Karpathy suggests a consistent prefix so the log can be read with grep, and so can you, at 11pm, wondering when you decided that.
3️⃣ Topic pages, made when they are needed and not before. Not a taxonomy. A taxonomy is a way of spending a Saturday.
Then the schema. Put this in ~/.claude/CLAUDE.md, which Claude Code loads in every project as your personal instructions:
That is the whole setup. Now the three operations are sentences you say. "Read this article and file what matters." "What did we decide about the database, and which page says so?" "Lint the brain." The first time it files something and updates the index unprompted, it feels like the future. The fortieth time, it feels like a colleague who takes good notes, which is better.
💡TIP
Keep the folder in git. Karpathy makes the same point: the wiki is "just a git repo of markdown files", so every change the model makes is a diff you can read and undo. A second brain that can quietly rewrite itself needs a history even more than a first one does.
📱 The catch: a brain in a jar
A folder on your laptop is a brain in a jar. Claude Code on that laptop can read it beautifully. Claude in a browser at work cannot. Claude on your phone, at the moment you actually remember the thing, cannot. A teammate, certainly not.
That is because Claude in the browser and on your phone reaches outside tools through connectors, and those connectors are servers on the internet, not folders on your desk. There are three honest ways around it.
Obsidian with Sync. This is the setup Karpathy describes, and it has a lot going for it. Obsidian is, in its own words, "Free without limits. No sign-up required." Your notes stay as files you own, the graph view is lovely, and Sync, at $4 a month billed annually, puts the folder on your phone. What Sync does not do is let Claude on your phone reach it. For that you also need an MCP server that can see the files, which brings back some of the steps you were trying to avoid. If you only ever talk to Claude Code on one machine, Obsidian is the right answer and you can stop reading here. Please do not.
A git repository. Version history, branches and backups for free, and a perfectly good home if you live in a terminal. On a phone it is a place you visit, not a place you write.
A notes app with its own MCP server. The brain lives at an address, so every Claude you use can reach the same one: Claude Code, Claude.ai, the phone app, and ChatGPT if you use that too. I compared eight notes apps that offer this, including who runs each server and what it can do, in a separate post. Notion's is the broadest. Obsidian's community servers keep everything on your machine. Pick by where you actually talk to Claude.
🧟 The part I build
This is the part I build, so adjust your trust accordingly. Sumibako is the third option: a notes app whose vault Claude can read, search and write through one connector address, from Claude.ai, Claude Code, ChatGPT and the command line. The schema above works with one change of address:
A few things are there because a brain that an AI maintains should have guard rails. The connection cannot delete pages. Every page it overwrites keeps its previous versions for two weeks, so a confident wrong rewrite can be put back. Pages are private unless you, or the agent at your request, publish one, which is also how you hand a single page of your brain to someone else without handing over the rest. The free plan holds a thousand pages, comfortably past the "hundreds of pages" Karpathy says an index alone can handle.
What it does not have: a graph view, and search that understands meaning rather than words. If you want to watch your brain as a constellation, Obsidian does that and I do not.
🧹 Keeping it alive
The way an LLM-kept brain goes wrong is not dramatic. A summary is slightly off, the next summary is built on it, and three months later a page states with total confidence something you never believed. Karpathy's answers are in the gist and they are the right ones: the raw sources stay untouched, so there is always something to check against; answers should cite the pages they came from; and the lint pass looks for exactly this drift. For a team wiki he suggests humans "in the loop reviewing updates". For one person, I read that as: skim the log on Fridays.
Do that, and the brain stays a brain. Skip it, and it becomes a diary of your life written by a well-meaning stranger, which is a genre, but not a useful one.
🦄
Is there a web version of Obsidian? No, and five ways around it
Obsidian has no web version. Five ways to reach a vault from a browser anyway, from GitHub's free editor to the real app on a server, and what each costs.
7 min read
markdown
## My second brainMy notes live in ~/brain. Read ~/brain/index.md before answeringquestions about my projects, decisions or past work.When we learn something worth keeping (a decision, a fact, a fix),file it: update the topic page or create one, add it to index.md,and append one line to log.md like "## [2026-10-03] decision | Title".Never delete a page. Mark a claim as superseded instead, with the date.When I say "lint the brain", look for contradictions, stale claims,pages missing from the index, and topics with no page.
markdown
## My second brainMy notes live in Sumibako. Search there before answering questionsabout my projects. File anything worth keeping with write_page, undera stable key like brain/<topic>, and append a line to brain/log.
