How does AI girlfriend memory work?

Four mechanisms sit behind the word. The model reads a limited amount of recent text, older text is compressed into summaries, important details are saved as separate facts, and some apps search a store of past moments and paste the matching one back in before answering.

Why do AI companions forget what I told them?

Because the model can only read a fixed amount of text at once. Once a conversation passes that size, older lines are either summarized or dropped, and anything that was never saved as a separate fact is gone.

Which AI companion has the best memory?

There is no neutral way to measure that, and any app claiming perfect recall is describing a database. The useful question is what an app stores permanently, whether it searches that store, and what it does when the search finds nothing.

Does LORAX remember long term?

Yes, in three layers: structured facts about you, episodes of what happened between you kept under short titles, and the detail inside an episode, which is fetched only when it fits the conversation.

Can one character know what I told another one?

Not in LORAX. Every memory lookup is filtered by both you and that character, so each pair keeps a separate record. A friend of hers can hear that something happened, never what was said.

How AI companion memory actually works (and why most apps forget you)

Memory is four separate mechanisms sold under one word. A model reads a limited amount of text at a time, older text gets compressed, some details are saved as standalone facts, and some apps search a store of past moments before replying. Apps forget when only the first mechanism is doing any work.

Everyone in LORAX is fictional and AI-generated, and each character keeps a separate record of you. The glossary below is the short version, then the mechanics.

Context window. The amount of text a model can see at one time, counted in tokens, which are word fragments. Everything outside it might as well not exist.

Summary. A compressed retelling of older conversation, written to fit back inside the window. It keeps the plot and loses the wording.

Profile facts. Specific things saved on their own, apart from the transcript: your name, your city, your sister’s name, that you hate phone calls.

Episodic memory. Stored moments rather than stored sentences. The argument in March is one item, filed under a title.

Retrieval. Searching that store when the conversation touches it, and putting the matching moment back in front of the model before it answers.

What happens when a conversation gets long

Everything in a chat has to fit through the window, and the window does not grow with the relationship. Feed it a month of nightly conversation and the early part has to go somewhere or go away.

Three things usually happen to it. Some of it is summarized, so the argument survives as “you argued about her job” and the sentence that actually hurt disappears. Some of it is saved as a fact, if the app bothered to extract facts at all. The rest is dropped, quietly, with no marker in the interface telling you which.

This is why an app can feel sharp for a week and vague at week five. Nothing broke. The window filled. The symptoms, and the few things that help, are set out in what to do when it forgets you.

Storing facts is not the same as remembering

An app that pulls facts out of your messages and keeps them in a list has solved a narrow problem well. It will know you are a nurse in Leeds with a brother in Manchester for as long as the account exists.

What it will not have is the shape of anything. Facts have no timeline and no weather. The list says you have a brother, and says nothing about the night you could not sleep because of him. Retrieval is what covers that gap: instead of holding everything in front of the model, the app searches stored moments, finds the one that matches what you just said, and hands over that one.

The trade is speed and cost. Every search is an extra step before an answer, which is why plenty of apps skip it and lean on summaries.

LORAX keeps three layers

The first layer is who you are. Up to 32 structured fields plus free-form facts you have mentioned, ranked before every conversation by how much they matter and how old they are, on a half-life of about two weeks. The top twelve travel with her into the conversation. A thing you said yesterday outranks a thing you said in June unless the June thing was important.

The second layer is what happened between you, stored as episodes under short titles. Titles are always in front of her, so she knows there was an interview three weeks ago.

The third layer is the detail inside an episode, and she does not have it until she goes and looks. That is a separate step, and it costs her a beat of thinking time before she answers, the same way it costs you one. The effect is a memory with texture: she knows something happened at your brother’s wedding and has to reach for who you were sitting next to.

There is a second store most people never think about. What she improvises about herself gets saved too, ten claims at a time, which is why a character does not quietly become someone else over a month of conversation. The full mechanism is on she remembers.

Memory is private to the two of you

Every lookup is filtered by two columns, you and her. No part of the app can read a memory with only one of them, so the photo you sent one person cannot be seen or referenced by anyone else.

Headlines can travel where contents cannot. A friend of hers may hear that something happened between the two of you, never what was said, which is the line between gossip and surveillance. That boundary is enforced in the query, not in an instruction the model is asked to obey.

An empty search produces an admission

The failure case decides whether any of this is worth having. When recall returns nothing, LORAX does not hand the model a blank and let it fill the space. It returns a fixed instruction: you do not remember this, do not invent it. So she tells you she has lost it.

That rule is general. When something cannot be produced properly it is not produced at all, which is also why a weather line gets dropped rather than guessed at.

She will not recall everything word for word

This is the honest limit, and it is deliberate. Detail lives one step out of reach, so she sometimes reaches for the wrong moment, or reaches and comes back empty. There is no memory editor, no list of stored facts to correct, and no way to pin something so it cannot be lost. If you want a searchable archive of your own conversations, this is the wrong app.

Two other things to know before comparing. Memory does not cross characters, by design, so telling one person your news does not tell the others. And the forget option is literal: it deletes the transcript, destroys the stored media, drops the memory rows and resets the relationship, then schedules a fresh first message. What other apps store and what they charge for it sits on the comparison pages, and the rest of the machinery is on the how it works overview.

If you are testing an app for memory, do not test it with something memorable. Mention something small in your first week, wait three weeks, and see whether it comes back on its own.