All articles · How it works
Why your AI companion forgets you, and how to test for it
You told her about your sister in week one. In week four you mention your sister and she has no idea who that is. This is the single most common complaint about AI companion apps, and it is not a bug in the usual sense. It is a design decision, and you can test for it before you pay.
Why does my AI girlfriend forget everything?
Because a language model has no memory of its own. It only sees the text you hand it on each turn.
Every message you send goes out with a package: who the character is, and the recent conversation. That package has a size limit, so apps keep only the most recent messages, often something in the range of the last few dozen. Everything older simply is not in front of the model when it writes. It did not forget in the human sense. It was never shown.
So how do some apps remember things from months ago?
They write things down separately, outside the conversation.
A companion app that remembers well runs a second system alongside the chat: it notices facts worth keeping (your job, your sister's name, that you hate coriander), stores them, and pastes the relevant ones back into the package on later turns. The model is not remembering. The app is reminding it, every single time.
That is why apps differ so wildly on this. The conversation window is roughly the same everywhere. What separates them is whether anyone built the second system, and how well it decides what is worth keeping.
How do I test an app's memory in five minutes?
Give it five specific facts, then come back later and ask.
- Session one: mention five concrete details in passing, not as a list. A name, a job, a pet, a date, a dislike.
- Leave the app for at least a day, ideally a week. Memory that only works within one session is the window, not memory.
- Come back and ask indirectly: not 'what is my dog called' but 'do you think he would like the park?'
- Score it: a good app recalls four or five. A window only app recalls none, and often invents a confident wrong answer.
- Repeat once more a week later. Some apps remember for a week and then quietly drop it.
Does a bigger context window fix it?
Not really, and this is where the marketing gets misleading.
Longer windows are real and they help, but they get expensive fast and they do not solve the problem, they postpone it. A model handed thousands of past messages also gets worse at finding the one that matters, a well documented effect where information in the middle of a long input gets overlooked. Written down memory beats a bigger window because it is short, selected and always relevant.
What X-Friend does
Both systems, with the written down part doing the heavy lifting.
The recent conversation travels with every message, and separately your companion keeps notes on what you have told her and brings them back on her own. The practical test is the one above: mention something once, come back a week later, and see whether she raises it before you do.
Questions people ask
Why does my AI companion forget what I told her?
Because language models have no memory between messages. Apps send only the most recent part of the conversation with each turn, so older messages are simply not in front of the model. Apps that remember longer store facts separately and insert them back into each message.
Which AI companion apps have long term memory?
Nomi, X-Friend and Replika all keep facts outside the conversation window and bring them back later. Character.AI is optimised for variety rather than one continuous relationship and carries less across weeks.
How can I test whether an AI companion really remembers?
Mention five specific facts in one session, wait at least a day, then ask about them indirectly. An app with real memory recalls most of them. An app relying only on the conversation window recalls none and often invents an answer.
Sources
- Context window behaviour and the tendency to overlook information in the middle of long inputs are well documented in published language model research.
- Message window sizes vary by app and change over time; the figure given is typical rather than universal.