How memory works in AI companion apps, and why most of it is worse than advertised

Every platform says its companion remembers you. They mean three different things.

The context window

The model reads a fixed amount of recent conversation. Anything older than that is invisible unless the platform does something extra. Cheap platforms stop here. The character forgets your name after a long session because it scrolled off the end.

Summaries

The platform periodically compresses old conversation into a summary and keeps that in the window. Better than nothing, and lossy. Specifics go first. The name of your dog survives. What you said about your week does not.

Long-term memory stores

The platform extracts facts and stores them separately, then retrieves the relevant ones each turn. This is what makes a companion feel continuous. It is what separates the top of our memory scores from the middle. Krush and Nomi do this best in our reading.

How to test it in a day

Tell the companion three specific, unusual facts in one session. Close the app. Come back the next day, chat about something else for twenty messages, then steer near one of the facts without naming it. A platform with real memory brings it up. A platform with a context window does not know what you are talking about. That is the 48-hour memory check in our testing protocol.