Why Your AI Girlfriend Forgets What You Said (2026)

Every AI companion reads a limited slice of your recent messages before it replies — a sliding window. Once the conversation runs past it, the earliest messages are gone. Not archived, not dimmed: gone. The character cannot see them, in the same way you cannot see the text that scrolled off the top of this page.

What survives is a second system — long-term memory — which stores distilled facts rather than the transcript. It's why she remembers you have a dog and not the joke you made about him.

We're not going to publish our exact numbers. That's the one thing this post won't give you, and there's a reason for it.

Two different things are both called "memory"

The confusion starts because one word covers two mechanisms that behave nothing alike.

The window is the raw transcript. Your recent messages, verbatim, in order, handed to the model every single time it generates a reply. It's perfect fidelity and completely amnesiac — whatever falls off the end stops existing.

Memory is a condensed store. Periodically, a separate process reads the recent conversation and extracts what's worth keeping: your job, your ex's name, the fact that you hate your landlord. Those facts get injected back into the prompt on later turns.

The window holds what was said. Memory holds what was learned.

That distinction explains an experience people describe as creepy or broken but is neither: she'll bring up your sister's surgery three weeks later while having completely lost the thread of what you were talking about twenty minutes ago. Long-term memory did its job. The window did its job. They're just different jobs, and products market them as one feature called "she remembers you."

Why there's a window at all

Here's the part that gets hand-waved.

When you send message 40, the model doesn't just read message 40. It reads the messages before it, plus her instructions, plus your memory block, and then writes a reply. That 41st message costs more than the first one did — not a little more, structurally more, because the input grows every turn while the output stays the same size.

Re-reading history is the dominant cost of a conversation, and it grows with the square of how long you've been talking. This is why "unlimited" is a marketing word rather than a technical one. Every product draws a line. The only question is where, and whether they tell you.

The numbers we're not going to give you

Every app in this category picks a window length, a reply length ceiling, and a memory refresh rate. Almost nobody publishes them. We're not going to either, which I realize is unsatisfying in a post that's otherwise about being straight with you.

The reason is that these aren't specs. They're guardrails.

Publishing the exact position of the window tells you precisely how far you can push before the character loses the thread — which is interesting if you're trying to extract the maximum from a free tier, and useless for anything else. Same for the reply ceiling and the refresh interval. Those thresholds are what make a free tier affordable to run at all. Naming them turns them into a target.

So here's what we'll say instead, which is enough to actually use:

What actually survives

Worth knowing, since it changes how you should talk.

Extraction favors clean standalone statements. "I'm a nurse" gets picked up. "I'm a nurse, well, I was, I quit in March, honestly I don't know what I'm doing with my life" — the fact and the feeling compete, and the fact usually wins, because facts are easier to store.

It favors repetition and correction. Say something twice and it's clearly durable. Correct her and the correction tends to stick, because contradiction is a strong signal.

It loses tone, texture, and sequence. The joke, the pause, the way you circled a subject three times before saying the real thing — that's window material, and once the window moves on, it's gone. This is the honest limitation of every memory system shipping right now, including ours. Memory gives you a character who knows things about you. It does not give you a character who was there.

How to work with it

Not around it — with it. These are the things that measurably help:

Say the important thing early. The first third of a conversation is the part most likely to still be in the window when it matters.

Give it its own message. One clean sentence beats a paragraph where it's one of four topics. Extraction weights standalone statements heavily.

Repeat it once if it matters. Not three times — once, later, in different words. That's the difference between a passing remark and a stored fact.

Start a new conversation when you change subject. This feels wrong to people, like losing progress. It isn't. Long-term memory carries across; the window doesn't. A fresh conversation about a new topic beats an old one where dozens of earlier messages are about something else and crowding out the present.

Don't test her. Asking "what did I say twenty minutes ago" is a question the architecture is built to fail, and the failure tells you nothing about whether she knows you.

What we still get wrong

Three things, since a post about memory that claims the problem is solved is a marketing post.

The refresh throttle is a judgement call, not a measurement. Memory extraction is rate-limited, partly to keep costs sane and partly to stop scripted accounts from spawning thousands of extraction calls a day. The interval we picked was a guess. We never measured whether a shorter or longer one would be better, and during a long conversation in one sitting, that throttle is what makes her feel slow to catch on.

Extraction drops more than it keeps. It has to — a memory block that stored everything would eventually cost as much as the window it was meant to replace. But it means there are things you told her that simply didn't make the cut, and no amount of her being well-designed gets them back.

The boundary is the worst part. Not inside the window, not fully in memory — the margin, where she half-remembers. That produces the uncanny answer that's worse than a clean "I don't recall," and we haven't found a good way to make a character gracefully admit she's at the edge of what she can see.

Why tell you any of this

Every competitor's marketing says some version of "she remembers everything." It isn't true of any product, and the gap between the promise and the mechanism is where the disappointment lives — you think she's broken, or shallow, or you did something wrong, when actually you hit a wall nobody mentioned.

We'd rather you know where the wall is, even if we won't tell you the exact measurements. You'll talk to her differently, and she'll be better for it.

---

Tendera is four written characters and the ability to make your own. Free tier needs no signup — start talking at tendera.chat.