Short answer
An AI second brain is a store of notes that an AI assistant can search, read and add to, so what you learn builds up instead of vanishing with each chat. The phrase covers three things: memory inside one assistant, AI built into one notes app, and notes you own that any assistant connects to over MCP.
An AI second brain is a set of notes that an AI assistant can search, read and add to, so the context you build up survives the end of a conversation. That much everyone agrees on. Where they part ways is the question that decides whether it is any use to a team: where the notes live, and who else can open them.
Three different products are sold under the name. One keeps the brain inside an assistant's memory, one keeps it inside a notes app with AI built in, and one keeps it in files you own that any assistant connects to. This page separates them, says which a team can share, and shows what the third kind looks like in practice, using Baalda, the open source team second brain this site is for, as the worked example.
What is an AI second brain?
The working definitions are close to each other. Tana's is the clearest: "a personal knowledge system that fills, connects, and updates itself from your work, then answers questions and turns thinking into filed tasks." Tana says it "captures from the work itself, your meetings, messages, and documents" and "updates existing records as new work arrives."
Strip the marketing and three jobs are left. Something captures what you learn, something stores it, and an assistant reads it back to you when it matters. Older second brains did the first two by hand. The AI version hands capture and retrieval to a model.
What the definitions skip is the store. The capture and the answers are the visible part. The store is the part you live with: it decides who can see the notes, who can correct them, and whether they survive the day you change assistant.
What are the three kinds of AI second brain?
| Kind | Where the notes live | Who can open them | When you change assistant |
|---|---|---|---|
| Memory inside an assistant | The assistant's account | One person | The memory stays with the old one |
| AI inside a notes app | That app's workspace | The workspace's members | The notes stay in the app |
| Notes you own, reached over MCP | Files you control | Whoever you share them with | You connect the next one |
Memory inside an assistant is what ChatGPT and Claude do when they remember you between chats. Anthropic's help centre describes Claude's version as something that "transforms from a stateless chat interface into a knowledgeable collaborator that builds understanding over time", and adds that "each project has its own separate memory space." It is useful, and it is personal by design. We cover the ChatGPT side in what ChatGPT memory is and what it is not.
AI inside a notes app is Tana, Notion and their neighbours: the notes and the model sit in one product. This is where the self-filling part is strongest today: Tana says it captures from your meetings, messages and documents as the work happens, and it runs an MCP server so other assistants can reach it too. Our comparison with Tana is straight about where it wins.
Notes you own, reached over MCP turns the arrangement around. The notes are plain files, and the assistant is a client that connects to them over the Model Context Protocol. Claude, ChatGPT, Cursor and Claude Code can all be pointed at the same store. The options for building one are compared in an MCP second brain is only as good as where its notes live.
Which kind of AI second brain can a team share?
Assistant memory cannot be shared, and it says so. Claude's help page on memory is explicit that memory is per person, and that even on a company plan "Owners can't view or edit a user's individual memories." Whatever your colleague's assistant has learned about the project, you cannot read it, correct it or inherit it when they leave.
A notes app with AI built in can be shared, inside that vendor's workspace. Tana writes that "a second brain does not have to stop at one person" and that "the team case is where it matters most", and for a team that is happy to live in one product, that is a fair offer. The trade is that the notes are in the vendor's store and the AI is the one the vendor chose.
Files reached over MCP can be shared and are not tied to one assistant. This is the kind Baalda builds: every note is a plain markdown file on your own disk, several people edit the same note at once over a CRDT, permissions are set per folder and per file, and the code is open source under Apache-2.0 and self-hostable. The assistant reads and writes those same files.
That last part changes what "the AI" in a team second brain is. It stops being one model in one account and becomes several connections, one per person and per tool, all writing into the same notes. Whether that is safe depends on whether you can see them.
What does the AI look like when a whole team connects?
In Baalda, each person creates their own connection under Vault settings, then MCP. A connection is a token that starts with mcp_ and is scoped to one person in one vault. The app shows it once and keeps only a sha256 hash of it, so the database cannot leak a working token. After you create one, the app hands you a ready-made command for Claude Code:
claude mcp add --transport http context https://api.baalda.com/api/mcp \
--header "Authorization: Bearer mcp_your_token_here"Every token then appears in a Connections list in the same settings page. Each row carries the name you gave it, the client on the other end (read from the request's User-Agent, so "Claude Code" or "Cursor"), the token's first few characters, how many tool calls it has made, and when it was last active. A dot marks it Connected if it has been used in the last three minutes. Next to it is a Revoke button, which deletes the row, and the token stops working on the next request.
Two properties make this hold up for a team rather than one person.
Each connection acts as the person who made it. A tool call from Claude Code is checked against that person's access, edit, view or none per file, exactly as if they had opened the note themselves. An assistant cannot reach a folder its owner was never given.
Membership is checked on every request. The server re-confirms that the token's owner still belongs to the vault each time the token is used. Remove someone from the vault and every assistant they connected loses access at once, without anyone hunting down tokens.
The call count only goes up on real tool calls, not on the handshake an assistant makes when it connects, so a busy number means an assistant is actually reading and writing notes. When an assistant writes, the change lands through the same CRDT path a person's typing does, so it appears live in any open editor and ends up in the .md file like any other edit. Setup for each client is in the MCP setup guide.
Where does a files-first AI second brain fall short?
It does not fill itself. Baalda does not record your meetings or read your Slack. Capture happens when you, or an assistant you have connected, write a note. If the self-filling part of Tana's definition is what you are buying, a meeting-capture product does that job better.
You bring the assistant. The model is whichever one you connect, on whatever plan you pay its provider for. Baalda stores and serves the notes. It does not include the assistant subscription.
Not every connection shows up in the list. Claude and ChatGPT can also connect through an OAuth sign-in instead of a pasted token. In the current source those connections are checked against the same per-file access and the same membership rule, but they have no token row, so they do not appear in Connections with a call count.
The labels are best effort. The client name comes from the User-Agent, so a script with a generic one shows up as "Node client" or under the first word of its User-Agent. Connected means used recently, not a live link, because MCP here is plain request and response over HTTP.
A solo second brain may not need any of this. For one person, assistant memory is the least effort, and a private scratchpad is better kept there than in a shared vault. The case for files reached over MCP starts when a second person, or a second assistant, needs to read what the first one wrote.
How do you choose an AI second brain?
Ask three questions before comparing features.
- Who needs to read what the AI writes? If the answer is only you, assistant memory is fine.
- Do you want the notes and the model from one vendor? If yes, a notes app with AI built in is the simplest route.
- Will you change assistant, or use several at once? If yes, put the notes in files you own and connect each assistant to them, and make sure you can see every connection and revoke any one of them.