Custody, not promises
Custody is the security model.
Your history is already spread across a dozen services you can’t search, can’t leave and don’t control. Second Brain Link gathers it on a machine you do: the build makes zero network calls, your exports are never uploaded, and what comes out is plain Markdown you own. Here is exactly what it does and doesn’t do — so you can check it against the source.
# common.py — the Collector every adapter writes through def add_person(self, source, name, …, email="", phone="", …): if not name or EMAIL_RE.search(name): return # never store a "person" whose name is an email # PII only enters in full mode; otherwise coerced to "" email = (email or "").strip() if self.full else "" phone = (phone or "").strip() if self.full else "" def add_message_signal(self, source, party_name, date="", ts=""): """Records ONLY that a message exchange happened + when. Never the body.""" # company/slack.py # NOTE: m.get("text") deliberately NOT read. # company/google_workspace.py QUARANTINE = {"payroll", "salaries", "loginaudit", "passwords", …}
Enforced in source you can read. View it on GitHub
What never leaves your machine
- The build has no network layer: detecting, matching and writing all happen locally.
- No telemetry, no analytics, no phoning home — run it fully offline.
- Where you point your AI afterwards is your choice; the brain is just local files.
What is dropped before anything is written
- Other people’s emails and phone numbers are rejected at parse time and never written.
- Message bodies are never read. Only who, with whom and when survives — a frequency and recency signal.
- Sensitive files are quarantined — passwords, login records, payment receipts and, for companies, payroll and security logs.
What you own at the end
- Plain Markdown in a standard Obsidian vault — no proprietary format. Delete it, fork it, grep it.
- Re-run any time from a fresh export. Want everything for your own brain? Owner mode (
--full) keeps it all — still local. - The agents’ memory lives in
_memory/in your brain: never uploaded, never used to train anything, and anything an agent only inferred is flagged for you to keep or remove.
scope“No telemetry, no analytics” describes the tool you download. This website uses consent-gated analytics — see the privacy policy.
Said plainly
What does touch the network.
“Zero network calls” is a claim about the build — the code that reads your exports and writes your brain. It is deliberately narrow, because a broad one would be false. Around it, these parts of the product use the network, each because you asked them to:
| Part | What goes out | When | To keep it local |
|---|---|---|---|
| Talking to your brain | Your question and the notes needed to answer it, to the AI you configured — Claude Code, Codex or a provider with your key — directly, with no server of ours in between. | When you ask | Run a local model (Ollama, LM Studio): nothing leaves your machine. |
| Agents | Requests to the public pages each agent declares in its manifest — job boards, a program’s own site, travel sites — read in your own browser. | When you run an agent | Agents are a separate install; they never run until you install and start them. |
| The Map | Google receives a Maps key and the area on screen to draw the tiles — never your notes. | When you open the Map | Don’t open the Map; everything else works without it. |
| The Avatar | The voice and hearing models download once — about 530 MB with a graphics card, about 170 MB without. After that, hearing and speech run on your device. | The first time you set it up | Don’t set up the Avatar. |
| Updates | The desktop app checks the public releases page for a new version. | In the background | The brain builder, the command line and the skill never check. |
| This website | Consent-gated Google Analytics — no analytics cookies until you accept. | When you browse secondbrainlink.com | Decline the banner. It is separate from the tool. |
Verify, don’t believe
Three checks, in ascending order of paranoia.
None of the above is worth anything as a promise — which is the point of shipping it as readable source.
- Pull the plug.Turn off Wi-Fi and build a brain. The build has no network layer to fail, so it completes offline. Anything that breaks was not the build.
- Watch the socket.Run the build under a firewall or connection monitor and count the outbound connections. The expected number is zero.
- Read the code.The rules are enforced where every source writes into the brain, in a language most engineers can audit in an afternoon. The repository
Custody scales better than policy. For a person, your life history sits on a disk you physically hold. For a company, it collapses a data-residency question from a jurisdiction to a room — and the security review to one answerable question: which of our machines is it on?
Not affiliated with or endorsed by LinkedIn, Meta, Google, Slack, X, Microsoft, Atlassian, Salesforce, HubSpot, Zendesk, Notion, Obsidian, Anthropic, OpenAI, the GBrain project, or any other platform whose exports it reads. It works with the data exports those services provide to you, and writes the open GBrain format.
Start
Your data already exists. It’s time it started working for you.
Free and open source — the privacy model included.
Second Brain Studio · Mac (Apple silicon and Intel), Windows and Linux · free · MIT
Just looking? Open Studio in your browser — two demo brains, nothing to install
