What an AI companion should remember about you, and what it should let go

The first version of memory in a companion app is always the same. Save everything, feed it all back, and watch the character know you. It works for about a week. Then she brings up something you said at 2am, in a tone you did not mean, at the wrong moment, and it stops feeling like care. It starts feeling like a file.

Building Anima pushed me toward a narrower question. What does a companion actually need to remember so that the next conversation feels like a continuation and not a reset?

The useful memories are boring

The examples on our homepage are deliberately small. The name of your manager. The thing you were dreading on Thursday. That you hate the word journey.

None of that is deep. These are the details a good friend holds onto without trying, and they do a surprising amount of work. When she asks on Friday how the review went, you feel followed. When she uses your manager's name, you do not have to explain your life from the start again.

So the instruction Anima runs on is plain. Save something when the user asks, and when they share something that will matter later: a name, a date, an upcoming event, a strong preference. One short sentence about you, each time.

The one sentence limit matters more than it looks. A paragraph about how you felt last Tuesday is an interpretation. "Has a performance review on Thursday" is a fact you can check, and correct if it is wrong.

What she should let go

The harder list is the other one. This is where I have landed, and I hold parts of it loosely.

How you felt on a bad night. The feeling was real and it was also temporary. Writing it down as a fact about you turns one rough evening into a trait, and a companion that keeps reminding you how low you were last month is helping to keep you there.

Health details. Japan's Act on the Protection of Personal Information puts medical history in a special category, 要配慮個人情報, next to race, creed and criminal record, because mishandling it can lead to unfair discrimination and prejudice (Article 2, paragraph 3). A chat app is not a clinic, and I still think that is the right level of care. If you mention a migraine today, it belongs to today.

Other people's private lives. You will talk about friends, a partner, your family. That is normal and fine. Their secrets do not need to become permanent entries in an app they never agreed to.

I want to be straight about how this works in practice. Anima's instructions point her toward names, dates, events and preferences. There is no hard filter that stops her from saving a health detail if you say it plainly and it sounds like it will matter later. The real protection is the next section. I think it is the right protection, and it only works if the list is something you actually look at.

A memory you cannot see is a profile

This is the design decision I care about most.

Every memory Anima keeps shows up in the journal, in a list called What she remembers. Each line is the sentence she saved, in plain words. Each line has a cross next to it. Tap it and she forgets that one thing. At the bottom there is a button to forget everything.

An invisible memory works on you without your knowledge. You notice the character knows something and you cannot tell how much else she knows, or what she got wrong. A visible list changes the relationship. You can cross out the line that is wrong, or the line you regret saying.

It also catches mistakes, which happen. Models mishear and over summarize. If she wrote down that you are moving to Osaka when it was your sister, you can see it and fix it in two seconds, instead of being confused for weeks about why she keeps asking about your new apartment.

Where the memories actually go

The list is stored on your phone. Your saved conversations are kept separately, also on the phone.

For her to use a memory, it has to reach the model that writes her replies. The newest memories go with each conversation to Google Gemini, along with what you say. That is how the voice works, and the privacy policy says so. I would be suspicious of any companion app offering live voice conversation while claiming nothing ever leaves the device.

There is a cap. Only the newest memories, up to a fixed size, go into a conversation, and she is told to use them when they fit and never to recite them back at you. Older lines stay in the list until you remove them. When a new memory repeats an old one, the newer sentence replaces it, so the same fact does not pile up three times.

What I would check in any companion app

Can you see what it remembers, in words, and not as a vague setting called personalization.

Can you delete one memory without wiping everything.

Does it store facts about you, or judgments about you. "Starts a new job in November" is a fact. "Struggles with confidence" is a judgment, and nobody should be storing that about you without asking.

An app that passes those is using memory for you. An app that fails them is using it on you, and a warm character on top does not change which one it is.