You tell a journal things you would never post. The fight you keep replaying. The doubt about your job. The person you miss. That’s the point of a journal: it’s the one place you don’t have to perform.

Now add AI. A journal you can talk to, that remembers and reflects back, is useful. But for an AI to read your words, those words have to go somewhere. For almost every AI app today, that somewhere is a company’s servers.

We built Merrin so they go nowhere. Its AI runs on your phone and your computer. Below is why, and what it costs you.

Is what you tell ChatGPT confidential?

What you say to a therapist, doctor or lawyer is protected by legal privilege. What you say to an AI chatbot is not. That isn’t our claim; it’s OpenAI’s CEO’s.

In July 2025, Sam Altman told Theo Von that if you share your most sensitive problems with ChatGPT and a lawsuit follows, OpenAI could be required to hand those chats over. His verdict: “I think that’s very screwed up.” (TechCrunch)

It wasn’t hypothetical. In May 2025, during the New York Times copyright case, a federal magistrate judge ordered OpenAI to preserve ChatGPT output logs that would otherwise have been deleted, including chats users had asked to delete (Engadget). OpenAI fought the order. It was lifted in October 2025, though logs already saved under it stayed accessible (Mashable).

To be fair to OpenAI, it didn’t want this, and Altman has argued for a legal “AI privilege”. But a court can overrule any company’s promise. Data that never leaves your device isn’t sitting on a server for anyone to order handed over.

Apple’s case for on-device AI

Apple made the same argument when it launched Apple Intelligence at WWDC 2024. Craig Federighi, Apple’s head of software, said you shouldn’t have to hand over the details of your life to be “warehoused and analyzed in someone’s AI cloud.” (Laptop Mag)

We agree, and we went a step further. Apple Intelligence still sends requests that need more computing power to Apple’s own servers, which it calls Private Cloud Compute (Yahoo News). That’s a careful design. Merrin has no AI server at all, so there’s nowhere to send a request.

Agentic AI and privacy

The next wave of AI doesn’t just answer. It reads your messages, your calendar and your files, and acts on them. Meredith Whittaker, president of Signal, warned at SXSW 2025 that this threatens to “break the blood-brain barrier between the application layer and the OS layer.” (TechCrunch)

Her point is that an assistant that sees everything becomes a single place where everything can leak, and much of that processing happens in the cloud. Merrin is built the other way round. Its AI lives inside one app, works on what you choose to tell it, and keeps it in an encrypted database on your device.

Why AI wants to be local

Pete Warden, a longtime deep learning engineer and co-creator of the Moonshine on-device speech models, has a four-word summary of where AI is heading: “AI wants to be local.” (The Data Exchange)

His reasoning is about shape. Google Search needs massive databases and armies of servers. What ChatGPT and Claude do, you can run on one laptop with no internet connection. He says local models stay only a few months behind the big commercial ones, and that the biggest barrier now is usability, not model quality.

We’ve found the same. Merrin’s standard model is a 2.6 GB file that runs on a phone with no connection, once the first 2.7 GB download (the model plus the voice models) is done. Pro adds a larger model, and Merrin Origin can run a larger one again on a Mac or Linux computer with 16 GB of memory. A phone model is smaller than the biggest cloud models, and the downsides of that are below.

Local AI has no per-message cost

When every request costs money, an app has to ration. Warden puts it this way: imagine app developers paying a few cents every time someone touched an icon. Apps would be designed very differently, and a lot of them wouldn’t exist. Going from a cent to zero, he says, isn’t an incremental change; it widens what’s possible. Apple pitched its on-device model to developers the same way, noting it runs without cloud API costs (Constellation Research).

That’s why Merrin can do things a cloud journal would struggle to afford:

  • No message limits, on the free plan too.
  • Work between conversations. When you’re not mid-conversation, Merrin pulls out what matters from what you said and saves it as memories and journal entries. With Pro, it also spots themes that come back across weeks and writes a weekly reflection. Each sentence of that reflection is checked against your notes from the week, and sentences they don’t support are dropped.
  • No ads, and no reason to want your data. There’s no server bill to pay off with it.

The downsides of local AI

Local isn’t free. The trade looks like this:

  1. A real download. Setup downloads the AI models once, a few gigabytes, over an internet connection. After that, Merrin works offline.
  2. Smaller models. A phone can’t run the largest cloud models. Merrin is built to listen, remember and reflect, not to be a general-knowledge chatbot.
  3. Your device does the work. Heavy tasks wait until you’re not in a conversation and run one at a time. On iPhone they run only while the app is open, and on a laptop they pause when the battery runs low. Some insights arrive later than a server would deliver them.
  4. Only you hold the key. Encrypted backups are protected by a recovery key we never see. Lose it, and we can’t get your data back. That’s the point, and it’s also the risk.

We think that’s a fair price for a journal no one else can read. You may weigh it differently, and that’s a reasonable choice.

How Merrin keeps your journal on your device

Merrin’s conversation model, speech recognition, voice and memory search all run on your device. There’s no account and no Merrin server for your content, so no server can read your conversations, journal, memories or photos. Everything is stored in an encrypted database, and you can lock the app with a PIN, a password or biometrics.

If you pair your phone with Merrin Origin on your computer, the two sync end-to-end encrypted, only between your own devices. On the same network they connect directly, with no Merrin server involved.

“Nothing leaves your device” would be too simple, so here’s exactly what does:

WhatWhenWhat anyone else can see
AI model downloads and configuration checksSetup, when you add a model, and when the app checks its configurationOrdinary network details, such as your IP address and which file was requested
Store purchasesWhen you buy or restore ProWhat Apple or Google see for any purchase. We receive the transaction details needed to unlock Pro.
Away-from-home sync (Pro)Only if you turn it onOur relay forwards encrypted data it can’t read. It sees IP addresses, a random ID for your pair of devices, traffic size and timing, and a proof of Pro. Never content.
DiagnosticsOnly when you turn them on and tap ShareTechnical details such as app version, device and errors. Never your conversations, journal, memories, voice or photos.
Phone speech recognitionOnly if you opt in, instead of Merrin’s on-device defaultApple or Google may process your voice
Exports and sharesWhen you choose to share somethingWhatever you choose to share

Test a private AI app in airplane mode

The easiest way to check a privacy claim is to cut the connection. Download Merrin, let the models finish, then turn on airplane mode and talk to it. Write, speak, ask it what it remembers. Everything still works, because nothing needed to leave.

Sources

Frequently asked questions

Does Merrin work offline?

Yes, once the AI models are on your device. Setup needs a connection to download them, 2.7 GB on a phone for the standard model and the voice models. After that, conversations, journaling, search and memory run on the device, so they keep working in airplane mode.

Can Merrin hand my journal over to anyone?

Merrin has no account and no server that stores your conversations, journal, memories or photos, so we have no copy to hand over. Your data lives in an encrypted database on your own devices, and in any backup files you choose to save.

Is a local AI model as capable as ChatGPT?

Not at everything. A phone can't run the largest cloud models, which know more about the world. Merrin's models are built for talking things through, remembering what you said, and reflecting it back, not for general-knowledge questions.

What does Merrin send over the internet?

Model downloads and configuration checks, store purchases, encrypted sync through our relay if you turn on Away from home with Pro, diagnostics only when you tap Share, and phone speech recognition only if you opt in. The table in this article lists what each one reveals. Your words never go to a Merrin server.