Until October 27, the first line of our privacy policy read: “Merrin is a private, local-first AI companion.” It now says “personal intelligence app”. Most of the site had moved away from the old word by then, and the policy was one of the last places it survived.

Swapping one word in a legal document sounds minor. We did it because “companion” describes a product whose job is to be liked, and to be missed when you put the phone down. We were building a place to think out loud that remembers what you said and helps you see your own thinking more clearly, and the old word kept pulling the product toward the wrong goal.

Saying that is easy. Keeping to it after launch, when growth targets start asking for something different, is harder, so the rules we’ve written into the app are listed further down, along with how to hold us to them.

What we mean by personal intelligence

Merrin starts with a conversation, spoken or typed, because most people think more clearly out loud than in front of an empty page. It turns what you said into a journal entry afterwards, so you don’t have to decide in advance that this is a journaling moment.

Memory is what makes that personal. Without it, every conversation starts from zero and you spend the first five minutes explaining your own life. With it, Merrin can recall what you’ve said before, connect a worry from this week to one from February, and use what it has gathered to make the next conversation more relevant.

The outcome we’re after is a clearer view of your own thinking: the decision you keep circling, the person who comes up whenever work gets hard, how your view of something has shifted since spring. General knowledge is a different job, and a large cloud model will beat a phone at it.

Where all of this runs, and why that matters, is in our private intelligence explainer.

What follows from that

The most useful thing Merrin can tell you about yourself is usually something you already said. An AI that gets to know you will be tempted to tell you who you are, and we think that’s the wrong job. Your own sentences, brought back when they matter, are more useful than a model’s theory about you, and much harder to argue with. Say you’ve told Merrin on three different evenings that you’re “fine, just tired” after calls with the same client. Pointing that out helps. Deciding what it says about your career isn’t Merrin’s call, and we’d rather it stay quiet than sound clever. A plausible guess about your motives is worse than no answer, because you might believe it.

An assistant that always agrees with you is a mirror, and mirrors flatter. In April 2025, OpenAI rolled back a ChatGPT update a few days after shipping it, because the model had become, in its words, “overly flattering or agreeable”. To its credit, OpenAI didn’t just revert it quietly. It published what happened, then a longer account of what it missed, and concluded that “personality and other behavioral issues should be launch blocking”. We agree, and the second post is worth reading for anyone building personal AI.

Two details in it matter for a product like Merrin. Several changes in that update, including a new reward signal from thumbs-up and thumbs-down ratings, had together weakened the signal that kept flattery in check, and OpenAI noted that user feedback “can sometimes favor more agreeable responses”. It also found that in some cases user memory made the effect worse, though it had no evidence that memory increases it broadly. For a product built on memory, that’s the most relevant line in either post. Any assistant tuned to in-the-moment approval can drift the same way, because being agreed with feels good.

That’s why Merrin has perspectives instead of one voice, and you choose which one you’re talking to. Listener gives you room to say what’s on your mind. Coach turns a loop into one next step. Devil’s Advocate makes the strongest case against what you’ve decided, one point at a time, and Future You weighs today’s choice against goals you mentioned months ago (those two are part of Pro).

Software that wants to be missed is working for someone other than you. If a product is sold as a friend, success looks like attachment: you come back often and stay long. We want you to close Merrin with a clearer head and not think about it until you need it.

Daily check-in reminders are off unless you turn them on. If you do, you get one a day at a time you pick, written without guilt or streak pressure, and none of them says “thinking of you”, because Merrin isn’t a person and shouldn’t talk like one. The app does show a card when you’ve journaled seven days in a row, but nothing warns you that you’re about to lose the run. We wrote about the harder version of this question, why anyone would open Merrin tomorrow, in the build story.

Privacy matters here for a practical reason before any principled one: people edit themselves when they think someone might read. You write differently in a notebook with a lock than in an email, and Merrin is only as useful as what you’re willing to tell it. If a thought has to travel to someone’s server before it can help you, plenty of people will say the polite version instead, and the memory fills up with that. Keeping conversations, the journal and memory on your phone is what makes the honest version possible.

For personal questions, context matters more than model size. The largest cloud models know far more about the world than anything that fits on a phone. But size doesn’t help with “should I leave this job?” if the model never heard what you said in March about wanting more time for your own projects. Merrin uses a smaller model on your device, and that’s a real trade: it’s weaker on knowledge-heavy questions, and our on-device vs cloud comparison is specific about where. For the conversations Merrin is for, we think it’s the right trade.

The rules we build by

Each of these is written into the app or the privacy policy today, so you can check it rather than take our word for it.

  1. The core of Merrin runs on your phone. Conversations, the journal, memory and search by meaning run on your device, and so does speech recognition unless you switch on your phone’s own recognizer. The network is used for things like downloading the AI models, purchases, configuration and optional diagnostics. If we ever add a feature that sends your words to an online AI service, the app will tell you before anything is sent and ask for permission where required. An ordinary conversation is never quietly handed to a cloud AI. (Privacy policy)
  2. No account. You don’t sign up, so there’s no account holding your conversations on our side. Deleting something in the app is immediate, and we don’t keep a server copy for you to chase.
  3. Memory you can see and change. You can view what Merrin remembers, edit or delete any memory, keep a conversation out of memory, or switch memory off. Incognito chats aren’t saved at all.
  4. It doesn’t make up your past. Every conversation runs under an instruction not to invent memories, personal facts, feelings, motives or past events, and not to claim to remember anything it hasn’t been given. A model can still get things wrong, so this is the rule it’s held to, not a guarantee. The memory controls are how you fix the misses.
  5. Observations need evidence. Merrin points out a pattern only when your own entries support it. Before it mentions a recurring theme, for example, the theme has to show up in journal entries on at least three different days within two weeks, and the note is written from those entries. The Analyst perspective is told to leave out diagnoses, unsupported causes and weak patterns.
  6. It doesn’t pretend to be a person. Merrin isn’t a therapist, doctor or emergency service. If someone may be in crisis, it’s instructed to respond with care and point them to emergency services or a crisis line.
  7. Your words aren’t the business model. We don’t sell your conversations, journal entries, memories or voice recordings, and we don’t use your journal or conversation content for targeted ads. Optional analytics, which ask for consent where it’s required, count events like app opens, not your text. Merrin makes money from an optional Pro plan.

None of this promises that Merrin is always right, that a phone matches a frontier model on general knowledge, or that your history survives a lost phone if you never made a backup. Those limits are in the downsides section of the explainer.

How to hold us to it

Most of the rules above are also in the privacy policy, which is the formal version. It carries an effective date, currently November 9, 2026, and material changes are announced in the app or another appropriate way where the law requires it.

This post is dated too. If we change one of these rules, we’ll say so here, with the date and the reason.

If you see Merrin doing something this page says it won’t, email us at [email protected].

Frequently asked questions

What is personal intelligence?

It's how we describe Merrin: an AI built around your own history rather than general knowledge. You talk things through, it remembers what matters from earlier conversations, and over time it helps you see patterns in your own thinking. In Merrin, that runs on your device.

Is Merrin an AI companion?

No. Merrin is a private personal intelligence. You talk or type, it remembers what matters from past conversations, turns them into journal entries, and points out patterns that your own entries support. It isn't designed to act as a friend, and it isn't a therapist, doctor, or emergency service.

Does Merrin send my conversations to a cloud AI?

Not during normal use. Merrin's core AI runs on your device, and your conversations, journal entries, and memories are intended to stay there. Some features use the internet, such as downloading the AI models and processing purchases. If Merrin ever offers an optional online AI feature, the app will tell you before your content is sent and ask for permission where required.

Does Merrin draw conclusions about me?

It's built not to. Merrin points out a pattern only when your own entries support it, for example a theme that comes up on several different days, and it doesn't diagnose you or guess at causes. It can still be wrong, so you can view, edit, or delete anything it remembers.

Can Merrin disagree with me?

Yes, if you choose a perspective built for it. Devil's Advocate makes the strongest case against a decision you've described, one point at a time, and Analyst points out themes and changes in what you've actually said. Listener, Coach, and Open Conversation are in the free plan; Analyst, Future You, and Devil's Advocate are part of Pro.