callaIf you have spent more than a week with an AI companion, you have probably had the moment. You mention your sister, the job interview, the dog that was sick, and a few days later the companion asks a question it should already know the answer to. The spell breaks. It was never really listening; it was just very good at replying.
I build Calla, an AI companion app, so I spend most of my time on exactly this problem. Here is what is going on under the hood, why most apps get it wrong, and what to look for if memory matters to you.
The context window is not a memory
Every chat model works from a context window: the slice of recent conversation it can see when it writes its next reply. Windows have grown a lot, but they are still a rolling buffer. When the conversation gets long enough, the oldest turns fall off the back. Close the app, come back tomorrow, and many apps start the model with an empty or nearly empty window.
So "it remembered what I said ten minutes ago" tells you almost nothing. Anything inside the window is easy. The question is what survives once it falls out.
The facts box is a start, not a solution
The common fix is a pinned facts box: a short list of things about you (name, age, job) pasted into every conversation. Some apps let you edit it; some fill it in automatically.
It helps, and it is also where most apps stop. A facts box holds a few dozen labels. It does not hold the story: that your manager has been difficult since the reorg, that you were nervous about Thursday, that your sister's wedding is in May and you are not sure about the speech. Those are the things that make a person feel known, and they do not fit in a list of labels.
Real memory is three separate problems
Doing this properly means solving three things, and each one can quietly fail.
1. Deciding what to keep
After each conversation, something has to read it back and decide what was worth remembering. Keep too little and you get the goldfish. Keep everything and the important parts drown in small talk.
The useful unit is usually an episode: a short summary of what happened and why it mattered, plus a few structured facts that changed.
2. Finding the right memory at the right moment
When you send a message, the system has to search everything it kept and pull back the few memories that matter right now. This is usually done with embeddings, numerical fingerprints of meaning, so that "my sister" can find the episode about the wedding speech even if you never say the word wedding.
This step is easy to get subtly wrong. A ranking mistake, such as sorting results so the least similar memories come first, still returns memories that look plausible, so nothing obviously breaks. The companion just seems a little vague about your life. We found and fixed exactly that bug in Calla in October 2026, which is why the advice to builders is blunt: store something specific, ask about it a week later, and check what the search actually returned.
3. Deciding when to bring it up
Remembering is not the same as caring. A companion that can answer "what is my sister's name" is not the same as one that asks, the day after, how the speech went.
That second behavior needs the system to track open loops, the things you said were coming up, and keep them open until they have actually happened. Many systems close the loop the moment the topic comes up again, so by the next day there is nothing left to follow up on.
A one-week test you can run on any app
You cannot see any of this machinery, but you can test it. If memory is the reason you want a companion, try this in the first week:
- Mention something specific with a date attached: an appointment, a deadline, a trip.
- Close the app. Come back the day after the date.
- See whether it knows about the thing at all, and whether it asks how it went.
The first part, knowing, is now reasonably common. The second part, asking at the right time without being prompted, is the hard one, and it is honestly where the whole category, Calla included, still has the most work to do.
Why it matters more than the pictures
Most companion apps compete on images, voices and how far the roleplay will go. Those matter. But what people describe when they talk about a companion that feels real is almost always memory: being remembered, not having to start over, being asked about the thing that was worrying you.
That is also where the trust question lives. Memory means the app keeps a record of what you told it. You should be able to see what it remembers, correct it and delete it, and the app should be clear about how that data is stored and who can see it. A companion that remembers you should not also be a liability.
The short version
If an AI companion forgets you, it is usually not because the model is not smart enough. It is because remembering is three separate engineering problems, keeping, finding and following up, and most products only solve the first one halfway. When you are choosing one, ignore the claims and run the test above. A week is enough to tell.