SecondBrain · Blog · AI second brain

AI second brain: what it is, three ways to build one, and which fits you

By Federico Cesarini, maker of SecondBrain · Updated November 4, 2026 · 12 min read

What is an AI second brain?

A classic second brain, in the sense Tiago Forte popularized, is a system you maintain by hand: you capture, organize, distill and express (his introduction). The AI version keeps the purpose and hands most of the work to a model. What stays yours is deciding what matters and checking what the AI did with it.

It is also not the same as the memory built into a chatbot. ChatGPT, Claude and Gemini each remember things about you, but that memory is the assistant's own working notes, shaped for its next answer. You can view and correct parts of it and copy some of it out, but it is not a full archive you can browse and move whole. If you want the idea itself, from commonplace books to CODE and PARA, start with our guide to the second brain. This one is about the AI kind.

Why AI second brains took off in 2026

Three things came together.

• Assistants got memory, each its own. Claude's memory is on by default on the Free, Pro, Max and Ultra plans (Anthropic), ChatGPT draws on saved memories and past chats (OpenAI), and Gemini can remember past chats (Google). None of them can read the others.
• A standard plug appeared. The Model Context Protocol, an open standard Anthropic introduced in November 2024, lets an assistant read from and write to an outside memory. Claude, ChatGPT and Gemini all accept custom MCP connections today, each within its own plan and region limits.
• Karpathy wrote down a pattern. On April 4, 2026 Andrej Karpathy published LLM Wiki, an idea file for a personal knowledge base that an AI agent writes and keeps up to date as linked Markdown pages, instead of rereading your documents from scratch for every question. Many "Claude Code plus Obsidian" second brain tutorials have followed.

The three ways to build an AI second brain

Almost every setup you will see is one of these three, or a mix.

1. The memory built into your assistant

This is the zero-effort route, and it is better than it was a year ago. Claude remembers your role, projects and the people and places in your work and life, and each Claude project keeps its own memory, knowledge base and instructions; projects are on every plan, five on Free (Anthropic). ChatGPT's memory, under Settings, Personalization, Memory, can draw on saved memories, past chats, files and connected apps, depending on your plan and region (OpenAI). On the web, Gemini keeps its memory under Settings & help, Personal context (Google).

The limits are the same everywhere. Each memory is locked inside one assistant. It is not a complete record: OpenAI says ChatGPT's memory summary does not show everything it remembers. And it is hard to browse, reorganize or move. Our post on using ChatGPT as a second brain shows how far this route goes.

2. A do-it-yourself LLM wiki with Claude Code and Obsidian

Karpathy's pattern has three layers. Your raw sources, which the AI reads but never edits. The wiki, a set of linked Markdown pages the AI writes and owns. And a schema, a file such as CLAUDE.md or AGENTS.md that tells the agent how the wiki is organized and what to do with a new source. Three routines keep it alive:

• Ingest: drop in a new article or transcript and the agent summarizes it, then updates every page it touches; Karpathy notes a single source might touch 10 to 15 pages.
• Query: ask a question and the agent answers from the wiki with citations. Good answers can be filed back as new pages, so your questions compound too.
• Lint: now and then, ask the agent to find contradictions, stale claims, orphan pages and gaps.

That is the difference from RAG, the retrieval most chat-with-your-files tools use: RAG digs fragments out of your documents each time you ask, while the wiki is written once and kept current. An index file lists every page so the agent knows where to look, and a log file records what changed and when. Obsidian is the usual window onto all of it; in Karpathy's words, "Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase." Claude Code fits naturally, since it reads CLAUDE.md files at the start of every session (Claude Code docs).

The strengths are real: plain files you own, full control over structure, and synthesis that improves as you add sources. The weak spot is the phone. The wiki lives at your desk, while most thoughts arrive on a walk, in the car or between meetings. You also own the upkeep of the schema and the cost of the agent's runs.

3. An app that captures and files for you

The third route moves the work to the moment of capture. You speak or share something, and the app writes the note, files it with related ones and turns dates into reminders, with no schema to maintain. The better ones also run an MCP server, so your assistants can read what you captured.

SecondBrain works this way. You tap and talk on iPhone, Android or the web; it writes a clean note, files it in the right folder and creates the calendar event or reminder you asked for. On the phone, photos and PDFs go in too. Later you ask instead of search. With SecondBrain Pro, the same memory is shared with Claude, ChatGPT and Gemini, within each assistant's plan and region limits, and on iPhone it exports Markdown that opens in Obsidian. Other apps such as Notion, Mem, Tana Outliner and Voicenotes now also offer AI and MCP access, several only on paid plans; our comparison of fifteen second brain apps covers them.

How to set up an AI second brain, step by step

Whichever route you pick, setting up an AI second brain comes down to five decisions: one home for your memory, capture that takes seconds, filing rules the AI follows, connections to the assistants you use, and a weekly check. Make them in this order and the system survives past the first week.

1. Pick one home for your memory. Your assistant's built-in memory, a folder of Markdown files, or a memory app such as SecondBrain. Two homes means two half memories.
2. Make capture take five seconds. A voice button on your phone, the share sheet, a quick note. If saving a thought takes longer than having it, the system starves.
3. Give the AI filing rules. A few folders for your projects and areas, which dates become reminders, what to ignore. In a do-it-yourself wiki these rules live in a schema file such as CLAUDE.md.
4. Connect the assistants you use. Plug the memory into Claude, ChatGPT or Gemini over the Model Context Protocol, so they answer from your notes instead of from scratch. Each vendor sets its own plan and region limits, compared below.
5. Ask instead of search, and review weekly. Ask your notes questions in plain language, correct what the AI misfiled, and spend ten minutes a week deleting what no longer matters.

With SecondBrain, steps 1 to 3 are the app itself: you talk, it files. Step 4 is a couple of minutes with SecondBrain Pro, and our shared memory guide has the steps for each assistant.

How to build an AI second brain with Claude

Claude can play three roles in a second brain: its built-in memory and projects for light use, Claude Code maintaining a Markdown wiki for a full do-it-yourself system, or a custom connector that plugs Claude into a memory you keep elsewhere. The three combine well: capture on your phone, then let Claude read it through a connector.

• Memory and projects. Memory is on by default on individual plans, under Settings, Your Claude. Give each area of your life a project with its own instructions and files.
• Claude Code and a Markdown wiki. Point Claude Code at a folder, write a CLAUDE.md that describes Karpathy's three layers and routines, and open the same folder in Obsidian to read what it builds.
• A custom connector. Every Claude plan can add a custom connector to a remote MCP server, one on Free; in the Claude app it sits under Customize, Connectors (Anthropic). That is how a memory you keep outside Claude becomes something it can search and add to during a chat. SecondBrain connects to Claude through its own plugin with SecondBrain Pro, and the shared memory guide has the current steps.

If you are coming from ChatGPT, Claude can also import a one-time copy of what ChatGPT remembers; our Claude memory import guide walks through it.

Can you build an AI second brain in Obsidian?

Yes. Obsidian keeps notes as plain Markdown files in a folder, so any AI agent that can read and write files can work on them, whether Claude Code, Codex or a community plugin. The most common setup is Karpathy's LLM wiki: the agent writes and links the pages, and you browse the graph in Obsidian.

The gap is capture away from your desk. One fix is to capture by voice in SecondBrain and, with SecondBrain Pro, let Claude Code read those notes over MCP, so what you said on a walk is there when the agent updates the wiki. On iPhone, SecondBrain Pro can also export your notes as Markdown that opens in Obsidian with no plugin.

Which AI is best for a second brain?

The best AI for a second brain is the one you already talk to every day, connected to a memory you control. Claude, ChatGPT and Gemini are all capable of the reading and writing. What differs is how much each remembers on its own and whether it can reach an outside memory on your plan and in your country.

Sources: Anthropic, OpenAI, ChatGPT pricing, ChatGPT import, Google custom apps and Google import, all read on November 4, 2026.

If you use more than one assistant, the second brain should live outside all of them. That is the case for a separate memory: you tell one assistant, and the others can look it up.

What an AI second brain still gets wrong

• It misfiles. AI puts some notes in the wrong place and occasionally gets a date wrong. Glance at anything important before you rely on it.
• Summaries lose detail. A compressed memory is not the original. Keep the source for anything you may need word for word.
• Privacy is a policy, not a feeling. Check where your memory is stored and whether it is used for training. Google, for example, says chats you import into Gemini are used to improve its services, including training models. SecondBrain stores notes in the EU and never uses them for training.
• Connections can be abused. An assistant that reads outside content can be steered by instructions hidden in it, which is called prompt injection. OpenAI warns that connecting untrusted MCP servers increases that risk, so connect only servers you trust.
• Lock-in creeps up. A memory you cannot export is a memory you rent. Plain Markdown files are the easiest to take with you.
• Do-it-yourself has running costs. A wiki that an agent maintains needs a schema you keep current and agent runs you pay for.

Questions people ask