Second Brain is on the App Store. It is a notes app built around one engineering claim: a roughly three-billion-parameter model, running entirely on your iPhone, can answer questions about your own notes, with citations, if you never let it guess.
That last clause is the whole product. Small models hallucinate. Ours is no exception, and we did not fix that. Instead we built a pipeline where a hallucination has nowhere to land: the model can only speak from retrieved notes, it must point at the notes it used, and when retrieval comes back empty it tells you it has nothing saved about that instead of improvising. In a memory app, a confident wrong answer about your own life is the failure that ends trust. So we optimized for refusing well.
From a raw thought to a cited answer
Here is the pipeline, in plain words.
Capture. You type a thought, speak it, or share it in from any app. It is saved raw, immediately, before any AI touches it. If the model fails at every later step, you still lose nothing.
Atoms. In the background, the on-device model splits each capture into small self-contained ideas (we call them atoms) and tags the people, companies, and projects they mention. “Dinner with Sara Friday, she’s allergic to peanuts” becomes two atoms, both tagged Sara. Atoms are the unit everything else operates on.
Embedding recall. When you ask a question, we embed it and score every atom by similarity, with a keyword boost on top, since embeddings alone miss exact names. A complex question is first split into up to three focused sub-questions, each searched separately.
Rerank. Embeddings are recall, not precision. They will happily return the wrong person with the right vibe: ask about April and get atoms about a different person from the same dinner. So the model runs a second pass as a strict relevance filter over the candidates: keep only the ones matching the question’s actual subject, or return none.
Grounded answer. The surviving notes are numbered and handed to the model with a narrow job: answer using only these notes, cite the ones you used, and if they are unrelated, output an abstain token. The app turns that token into a plain “nothing saved about that.” If the first pass finds nothing, the question is rephrased once with broader words and retried, because people rarely ask with the words they wrote.
The judge we deleted
The original design had one more model stage at the end: an LLM “relevance judge” that read the finished answer and decided whether it really answered the question. It sounds right, and it works in papers. Measured on-device, on our own eval set, it rejected correct answers and accepted wrong ones.
So we deleted it. Everything after synthesis is now deterministic code: citations must point at real retrieved notes, the abstain token bypasses everything, and empty retrieval never reaches the model at all. Each model call in the pipeline has one narrow generative job; the trust decisions live in code we can test. It is the same lesson we wrote up in a weak model and a strong harness: put the model where it is strong, and never let it grade its own work.
A privacy model you can check with airplane mode
There is no Studio AM account or AI-processing server. Turn on airplane mode after the on-device model is ready and capture, organizing, Ask, and local reminders continue to work. “Remind me Thursday” inside a captured thought becomes a real reminder on Thursday, scheduled on the device. Notes stay local unless you enable optional private iCloud sync, which is off by default and inaccessible to Studio AM; voice-recording audio remains local.
This was a constraint before it was a feature. A memory app collects the least shareable text a person produces: half-formed decisions, things about other people, things you are still working out. We compared how the main ask-your-notes apps handle that data, and almost all of them process it on a server, in readable form, at the moment you ask. We wanted the version where the question never comes up.
On-device also should not mean locked in. Note and Ask text exports to structured JSON or Obsidian-friendly Markdown, and text or Markdown files import back in. Recording audio and photos are not included, so these exports are portable text records rather than complete or restorable backups. Optional iCloud sync, off by default, can keep supported note data in your own private iCloud.
Supported devices can use a downloadable model
On Apple Intelligence devices, the built-in tier runs on Apple’s foundation model through the Foundation Models framework.
Other supported iPhones and iPads may be offered a one-time downloadable model during setup. Once it is ready, it runs on-device and offline with the same citation and abstention rules. Unsupported hardware cannot run the AI features, and the app explains the available path rather than promising support from a model name or device-age shortcut. The harness does not care which supported model sits inside it, and we run our evaluations against each tier separately. A prompt that works on one small model routinely breaks on another, and the evaluation suite is what catches it.
Where it falls short
We would rather you know these before downloading.
- English is the strongest language. You can capture in other languages, but recall quality drops. We are improving this tier by tier.
- The built-in tier needs recent hardware. Most other recent iPhones (roughly iPhone 11 and later) pay in storage instead, after a one-time download (2.6GB for Gemma or 0.9GB for the lite model), while the oldest, lowest-memory iPhones can’t run any on-device model.
- It only knows what you captured. It is not a general assistant and will not answer from the web. An empty brain says so.
- Answers are short, plain sentences. A small model synthesizing across your notes writes like a careful assistant, not an essayist. Every answer is AI-generated, so review anything important.
What the subscription covers now
Second Brain no longer has a functional free plan. Eligible new subscribers may receive a 3-day trial; after that, the US price is $29.99 a year or $4.99 a month. An active subscription authorizes new notes, recordings, imports, reminders, Ask, Reflect, and other AI output.
The archive remains yours when a subscription ends. Previously saved notes stay readable and searchable, voice recordings stay playable, note and Ask text remains exportable, and existing items can still be deleted. The retired lifetime unlock is not offered to new customers, but verified previous owners keep full access and can restore it through the App Store.
Second Brain is on the App Store today; details, screenshots, and requirements are on its page. If you get it to hallucinate, tell us exactly how. The abstain rate is the number we care about most.