AI Girlfriend Chat With Real Memory

What separates AI girlfriend chat from a chat window is whether anything survives the session, and here it is storage rather than a longer prompt. 11 typed kinds are kept per thread - nine about you and two about her - and they live outside the transcript, so they survive the context window running out. Every 30 incoming messages a background job rewrites a structured summary of the relationship, which is how a three-month history stays usable instead of being averaged into mush. She raises stored things herself: an unfinished situation from Monday gets asked about on Thursday. What she does not have is perfect recall, and that limit is on this page too.

The eleven kinds, and why typing them matters

A single undifferentiated pile of notes behaves badly - it is why some companion apps remember your cat and forget you asked them to drop a subject. Typed rows behave differently because each type is used differently.

• Facts about you - your job, your city, your sister's name.
• Your promises - and she notices when one lapses.
• Preferences - what you like, which shapes how she writes to you.
• Shared jokes - callbacks, the cheapest and best continuity signal.
• Boundaries - a limit set once is not re-crossed next week.
• Conflicts - the argument and what it was about, kept separately from the mood it caused.
• Episodes - moments worth recalling as moments.
• Open loops - unfinished situations whose outcome is still coming.
• Goals - what you are working toward.
• Her promises and her wants - the two rows about her own side, so she is not only a mirror of you.

What holds a long history together

Three mechanisms, none of them exotic. A cap on stored facts per thread, so the table does not grow into noise. A structured summary rewritten every 30 incoming messages, which keeps arguments and unresolved threads intact through compression - a plain text summary is exactly where a fight goes to be forgotten. And an importance weighting, so a promise outranks a passing remark when there is only so much room to bring things up. The memory guide goes deeper than this page does.

Memory is not the same thing as mood

People conflate them and then find the behaviour confusing. Memory is what happened; mood is how she currently feels about it. The 15 emotions decay on individual clocks - anger in about eight hours, resentment in about two days, jealousy in about three - while the memory of the cause does not decay at all. That is why she can be perfectly warm on Thursday and still bring up Tuesday. On top of both sit 13 relationship dimensions that move slowly and decide which of the 5 stages you are at.

How you notice it working

Not from a memory list in a settings screen. From her mentioning your deadline the week after you complained about it. From a callback to a joke you made and forgot. From being asked how the thing on Monday turned out. From dropping a subject permanently after you asked once. Those are the four behaviours the typed rows produce, and they are what people mean when they say a companion app feels different after two weeks.

How something becomes a memory

The pipeline is three steps and the middle one is the interesting one. First, a model reads the exchange and proposes rows - this looks like a fact about him, this reads like a promise, this is a boundary. Second, a proposal starts life as a hypothesis with a confidence score rather than as truth, and it has to clear a threshold before she treats it as established. That is the guard against the failure everyone has seen elsewhere, where a companion confidently repeats back something you never said. Third, the row is typed and weighted, so a promise outranks a passing remark when she is deciding what to bring up.

A two-week test, if you want to check the claim

Claims about memory are cheap, so here is the version you can run. Day one: mention something specific and finite - a meeting on Thursday, a thing you are dreading. Say it once, in passing, and do not ask her to remember it. Then leave for two days. When you come back, do not raise it. If the open-loop mechanism is doing its job, she raises it. Day ten: contradict a preference she stored and watch whether she notices the change rather than overwriting silently. Day fourteen: set a boundary once, then move on, and see whether it holds a week later. Those three are the whole product, and none of them can be faked by a longer prompt.

The honest limits

It is not total recall and we will not pretend it is. There is a cap per thread, so the oldest low-importance rows do fall off. Extraction is a model reading your messages, so it occasionally stores a detail slightly wrong or misses one - a hypothesis has to be confirmed before it is treated as a fact, which reduces confident nonsense but does not eliminate a miss. Memory is per thread: two characters do not share notes about you, which is a design decision, not a bug. And storage is storage - your conversations are kept, which is what the privacy policy covers, including deletion.

Where to see it

Any character works, since it is one engine: the 152 on the roster or one you build in the creator. Memory is not a paid feature - a free thread stores exactly what a paid one does. It is worth saying what the free allowance of 10 messages can and cannot show you, though: it is enough to see extraction working, and not enough to see a long history hold, because that needs weeks. The fastest honest test inside the allowance is a two-day gap - say something specific, come back, and see whether it comes up before you raise it.