We make one of the two things on this page. Brownie is ours, and Apple Intelligence ships free with the operating system we depend on, so read this the way you would read any comparison written by the smaller party. Every claim about Apple below comes from Apple's own newsroom posts and machine-learning research, linked where it is made and checked on 1 November 2026. Where Apple has not said something, we say that it has not said it rather than filling the gap with a guess.
You ask; it searches; it answers. Apple Intelligence is very good at this, and on the Mac it is now excellent — Siri AI searches your messages, mail and photos, answers from what it finds, and then acts on the answer. When the question is already formed in your head, this is the right shape of tool, and it is installed.
Something that read the last week without being asked and wrote down what it found, so the things you would never have thought to ask about are already on paper. Nobody has to form a question. This is the half almost nobody builds, and it is the only reason Brownie exists.
The distinction matters because the two fail in completely unrelated ways. Retrieval fails when you ask badly, and the fix is a better question. A record fails when it quietly stops being true, and the fix is maintenance — every night, forever. That is a far less interesting engineering problem, and it is most of what Brownie actually is.
It would be easy, and wrong, to say that Apple Intelligence only ever reacts. Safari's Notify Me watches a page and tells you when it changes. Passwords will navigate to a site on its own and replace a weak password for you — genuinely agentic, and further than Brownie goes, because it presses the button. Mail and Messages offer one-tap suggestions you did not ask for. Siri Recap writes a title and key points after a conversation ends, and Health Insights updates a summary through the day.
So the claim worth testing is narrower than “it does nothing in the background”. It is this: none of it accumulates. Each of those features does its job and finishes. There is one durable store — the Spotlight index, which Apple says works entirely on device — but an index exists to be searched, not to be read. Nothing Apple has published describes a thing that holds what matters about a person and keeps on holding it.
That is not a gap Apple failed to close. It is a different product, and reading Apple's own sentence closely is the fastest way to see it: find what they need in the moment. Every word of that is doing work.
| Apple Intelligence & Siri AI | Brownie | |
|---|---|---|
| What it is | An assistant built into macOS, plus writing and photo tools | A reader that runs while you are asleep and writes notes |
| How work starts | You ask, Control-click, or tap a suggestion | On a schedule you set. The Mac wakes itself, lid shut, on AC power |
| What it keeps | Siri's conversation history, synced over iCloud. Apple documents no maintained record of people | One Markdown note per person, kept for years, with what is open between you in both directions |
| Where it runs | AFM 3 Core and the 20-billion-parameter sparse AFM 3 Core Advanced on device; larger requests on Private Cloud Compute | A 4B model on your own GPU does the reading. Summaries go to a cloud model you chose — or nowhere at all |
| Personal context from | Apple names messages, email and photos | iMessage, WhatsApp, Telegram, IMAP mail, Slack, Teams, Apple Notes, Calendar, files, meeting recordings, MCP |
| Third-party apps | Acts in them: WhatsApp and Audible today; Outlook, Notability and Tripsy announced | Reads them, and acts in them |
| Acting for you | Completes the action | Fills the draft and stops before Send. Enforced in code, not in a prompt |
| Requirements | macOS 27, Mac with M1 or later. Siri AI is in beta, English first | macOS 14 or later, Apple silicon, 8 GB. A 3.7 GB model download, once |
| Limits & price | Included with macOS. Server-side features carry daily usage limits, and Apple says expanded access “will be available for a fee in the future” | Free, AGPL-3.0, no account and no server |
| Checkable | Apple publishes its model research and the Private Cloud Compute design. The code is not open | Source is AGPL-3.0, and the app lists every byte that left the Mac for thirty days |
Checked 1 November 2026 against Apple's own announcements: the June 2026 Apple Intelligence release, the September 2026 Siri AI release and Apple's third-generation foundation models. Apple ships quickly and some of this will age. Correct anything by opening an issue.
The best thing in Siri AI is that you can Control-click an image, a file or a block of text on your Mac and ask about it. Apple's words: “With onscreen awareness, Siri AI can answer questions or take actions related to the content on a user's screen.” It works, it is a lovely interaction, and it is exactly scoped to what is visible.
Which is the inverse of the problem most people have. The message that costs you something is never the one on screen. It is the one from three weeks ago, in a thread you have scrolled past two hundred times since, where you said you would send someone the numbers.
Brownie writes one Markdown file per person and keeps it for years. Under the title sits a ledger of what is open between you, in both directions: ⏳ for an ask or a promise outstanding, ✅ once it is answered, ⌛ when it has been let go — an ask after forty-five days, a promise after ninety. Code writes that block, not the model, so it cannot drift into something that merely sounds right.
The rest of the note has a shape, and the shape is enforced every night: standing facts that need no date, at most six dated bullets for where things stand now and none of them older than forty-five days, older context beneath that, then one line per month going back a year and no further. A gardener pass rewrites it so a year of nightly runs cannot bloat it into something nobody will open. A person who goes quiet for six months moves to an archive, and comes back the moment they are mentioned again.
None of that is exciting, and all of it is the point. It is what decides whether the thing is still worth reading in March.
Apple's position is strong and honestly engineered. The on-device models run on your Mac, and anything larger goes to Private Cloud Compute, whose design Apple has published and which Apple says holds no personal data once a request is done. At that scale there is very little else as well built. What it asks of you is trust: the guarantee is excellent, and the code implementing it is not yours to read.
Brownie asks for something different and hands back something different. Raw messages, mail bodies and files never leave the machine — not as a policy but by construction, because the only thing the pipeline can emit is a summary that has passed a deterministic PII backstop. What does leave is that short summary, to whichever model you chose with your own key, and the app lists every byte of it for thirty days under What left your Mac. Choose “This Mac only” and the figure is zero: no key, no account, no server, nothing to sign out of. An item flagged sensitive leaves no title, no summary, no card and no line in Brownie's own logs — a count, a time and a source, nothing more.
Neither answer is right for everyone. Extending trust to a very well-built system is a perfectly rational thing to do, and most people should. The difference is only in what you are holding afterwards: a guarantee, or a receipt.
These two do not compete for the same slot. One is the assistant you ask; the other is the thing that noticed. Running both is the sensible configuration, and they compose better than you might expect: Brownie writes plain Markdown into a folder in your home directory, so Spotlight indexes those notes alongside everything else and the record becomes searchable by the tools you already use.
At an hour you choose, your Mac wakes itself and a four-billion-parameter model on your own GPU reads everything new since the last run — messages, mail, notes, files. What matters goes into those notes. In the morning there are a handful of cards: a reply someone is waiting on, a promise you made, a renewal worth cancelling, each with its evidence and a draft in your own words.
The same split turns up elsewhere, and we wrote those down too: local AI assistants for Mac — nearly all of them are a chat window, very little reads your Mac. AI and Obsidian — every plugin reads the vault, almost nothing fills it. And on the thing these notes are for: what did I promise?
It is free, AGPL-3.0 and in the open, so everything above is checkable rather than promised.