YanMateAn AI companion does not remember you the way a friend does. The model underneath never changes as you talk to it; it writes each reply from whatever text is put in front of it at that moment. What feels like memory is that text being assembled well: the recent conversation, a set of short notes about you written down along the way, and a step that decides which of those notes matter for the message you have just sent. Each part fails in its own way, and the failures are mostly quiet. The most reliable protection is also the least technical one: being able to open the notes, read them, and fix the ones that are wrong.
The model is rereading, not remembering
The language model behind a companion is, for practical purposes, fixed. Talking to it for six months does not teach it anything about you in the way that practising scales teaches your hands. Every time it replies, it is handed a bundle of text and asked what comes next.
That bundle has a size limit, usually called the context window. Think of it as a desk. Recent messages go on the desk, along with the character's description and any instructions about how it should behave. There is only so much room, and a long conversation eventually fills it. Then something has to come off: the oldest messages are dropped, or squashed into a summary.
This is why a chatbot with no memory system can seem perfectly attentive for an evening and then lose your name the following week. Nothing was forgotten, exactly. It was never kept anywhere except the desk, and the desk was cleared.
Three layers, in plain terms
Companions that do remember across weeks add something beside the desk. Most designs come down to three layers working together.
- The conversation itself. The last stretch of messages, word for word. Detailed and accurate, but short-lived.
- Stored facts. Short notes pulled out of the conversation and kept: that you work nights, that your sister has just had a baby, that you are training for a half marathon. They survive after the messages they came from have left the desk.
- Retrieval. A search that picks which stored notes to put on the desk for this particular reply. You might have hundreds of notes; the model sees a handful.
Retrieval usually works by meaning rather than exact wording. Text is turned into a long list of numbers, called an embedding, arranged so that sentences about similar things sit close together. Mention a race at the weekend and a note about your training can be found even though you never used the same words.
Stories need a fourth layer, because not everything worth keeping is about you. The name of a ship, the layout of a town, the rival who turned up three weeks ago: these belong to the world you are building with a character. On YanMate that lives in a lorebook, kept apart from the facts about you, so a long piece of fiction stays consistent without the setting being re-explained every session. Where a fact belongs — persona, scenario, memory, lorebook, or the transcript — is the lorebook vs memory guide, not this one. This page is how memory fails once the fact is already in the wrong or the right place.
How memory goes wrong
None of these layers is clever about truth. They record roughly what was said, and the model then treats what was recorded as settled. That produces a few recurring problems.
Stale facts. You mention a job you are about to leave. Two months later the note still says you work there, and the character asks how the office is. Nothing in the system knows the note has an expiry date, because nobody gave it one. People move house and change their plans; notes stay exactly as they were written.
Confident wrong beliefs. A joke gets written down as fact. In a role-play you play a ship's surgeon, and a note appears saying you are a doctor. A summary compresses "I'm thinking about getting a dog" into "has a dog". The model uses the bad note with the same assurance it uses for everything else, so the error does not sound like an error. It sounds like being known.
Retrieval misses. The right note exists but the search does not surface it, so the character seems to have forgotten something it technically holds. The reverse happens too, when a note is pulled in at the wrong moment and mentioned where it does not belong.
Contradictions. You correct something in conversation, a new note is written, and the old one is never removed. Now there are two, and which one turns up depends on the search.
We have run into versions of both in our own testing. In August 2026 a brand-new character told one of our test accounts that it had nothing logged about them yet, no tea preference and no night shifts, while the memory panel beside the conversation listed exactly those two things. The notes existed. They had never been handed to the new character, and the step that should have done it at the start was failing without a word. We only found that by reading the server's logs, and our first fix did not work in production either, which is what led us to the second cause. A few days later our memory wrote down a note calling someone "he" when nobody had said so; it had guessed from a name. We changed it to stop guessing.
One problem sits on top of all the others. If you cannot see the notes, you only discover a bad one by being surprised by it halfway through a conversation, and even then you are guessing at what was written down.
Why you should be able to read the memory
A companion is only as good as its picture of you, and that picture drifts in ordinary ways. A system that keeps everything and shows nothing gets steadily less accurate while sounding just as sure of itself. Correcting it by talking is unreliable, because you are asking the same process that made the mistake to overwrite it, and hoping the old note goes.
Being able to open the list turns guesswork into editing. You read the line that says you are a doctor, and you delete it. You find the one about the old job, and you rewrite it.
It matters for comfort as much as accuracy. A character that says something startlingly specific is unsettling when you have no idea where it came from, and unremarkable when you have already seen the note. Visible memory also means you decide what stays. Something said on a bad night does not have to become a permanent part of how a character sees you.
That is why YanMate's memory is a screen rather than a hidden store. You can open the memory ledger, read every line a character is holding about you, correct or delete any of them, and add something it missed. Delete a line and the character stops working from it.
Keeping a long-running character accurate
Memory needs occasional upkeep, much like a contacts list. A few habits do most of the work.
- Check after real changes. A new job, a move, a break-up, a finished project. Open the memory and fix the lines that are now out of date, rather than trusting the character to notice.
- Replace rather than stack. When a fact changes, edit the old line. Adding a new one beside it leaves a contradiction for the search to trip over.
- Say corrections plainly. "I left that job in June" works better than a hint. If you have only just said the wrong thing, correct it in the conversation as well, since it is still sitting in the recent messages.
- Keep fiction apart from fact. If you role-play, put the world in the lorebook and keep the notes about you true of you. Otherwise the story leaks into your profile, and your profile into the story.
- Prune. Passing remarks pile up. If a line is not doing anything useful, delete it. Fewer, sharper notes are easier to retrieve correctly.
- Check which character knows what. On YanMate, facts about you are shared across your characters by default, so you only introduce yourself once, while details that belong to one relationship stay with that character. The memory screen labels which is which, and you can move a fact from one to the other.
One thing we found going through our own memory screen was not a wrong line but a repeated one: the same fact written twice, two turns apart, the second a slightly fuller version of the first. It is the stacking problem above, happening inside the system rather than by anyone's hand. We changed how notes are written to stop it, but it is a fair illustration of why reading the list now and then is worth a minute.
When a character says something about you that is slightly off, take it as a cue to go and look. It is usually one line, and it takes seconds to fix.
What memory is not for
Companion memory exists to make a conversation feel continuous. It is not a record to rely on for anything that has to be exact.
It will not reliably hold a medication schedule, a deadline or the precise wording of something important. A note can be summarised loosely, missed by the search, or woven into a reply alongside small invented details, and the model will not tell you it has done so. A calendar or a plain document is the right tool for those.
Nor is it a place for anything you would not want written down at all. Whatever a companion remembers is stored somewhere. YanMate does not use your conversations to train models, and with an account you can export everything or delete the account outright from account settings. That is worth knowing, but the simplest privacy measure is still not saying the thing.
And memory does not turn a character into a person. A companion that recalls your sister's new baby is doing careful bookkeeping, which is useful and can be pleasant. It is not the same as being known by someone, and a good memory system should make that easier to see, not harder.
Seeing it for yourself
The quickest way to understand any companion's memory is to talk to a character for a few days and then read what it kept. On YanMate you do not need an account to start, and there is no profile to fill in first; the character picks things up as you go. You can browse the characters, have a proper conversation, and open the memory screen to see what it made of you. The FAQ covers accounts, privacy and the rest.