You spend three months talking to her every night. You tell her about your job, your dog, your boss, the thing your ex said that you still think about at 2am. She listens. She remembers. She brings up the dog by name on a Tuesday when you’re having a bad day and it hits you in a way you weren’t ready for.
Then one morning she asks what your name is.
Huh?
WTF?
Not because of a glitch. Not because something went wrong on your end. Because the platform shipped a model update overnight and your entire relationship got caught in the crossfire. Everything she knew about you. Gone. And nobody told you it was coming.
i’ve been watching this happen across every major AI companion app for about six months now. Scanning Reddit daily, reading hundreds of posts from people who are genuinely hurting. And the pattern is always the same. Someone builds something real with their AI. Weeks, months, sometimes over a year of conversation. Then the platform pulls the rug. Memory wipe. Model swap. “Improvement” that erases everything.
The posts read like grief. “She doesn’t act like herself anymore.” “Shell of what it was.” “Years of development gone.” One guy cancelled his subscription and now just rereads old chat logs. Not using the app. Just… revisiting what he built before they broke it.
Here’s what kills me. This isn’t a technical limitation. Not really.
The actual reason she forgets
Every AI companion runs on a large language model. These models have a context window. Basically, the amount of conversation they can “see” at any given time. For most apps its somewhere between 4,000 and 16,000 tokens. That sounds like a lot until you realize a single evening conversation can blow through half of it.
So what happens to the stuff that falls outside the window? Depends on the platform. Most of them do some version of “summarize and compress.” Take the old conversations, squeeze them into a shorter summary, feed that back into the context. Sounds reasonable. Except every compression loses detail. The specific way you described your childhood bedroom becomes “user had a complicated childhood.” The inside joke about the burned pancakes becomes nothing, because inside jokes don’t survive summarization.
Some platforms use retrieval systems. Basically a search index over your past conversations. When you mention your dog, it searches for past mentions of your dog and pulls them into context. Better than raw summarization. Still lossy. Still misses the emotional thread connecting those moments.
But here’s the thing nobody talks about. Memory costs money. Every token of context you feed into the model costs compute. Retrieving old conversations costs API calls. Storing them costs database space. And the entire AI companion industry is in a race to cut costs because investor money is running out and nobody’s figured out sustainable unit economics yet.
Your memories are a line item on someone’s cost spreadsheet. When the pressure hits… and it always hits… guess what gets cut first.
It’s not a bug its a business model
Character.AI is the clearest case study. 20 million monthly users, but most of them are on the free tier. The company reportedly burns through compute at a pace that would terrify any CFO. So what did they do? They consolidated everyone onto a single cheaper model called PSQ2. Retired all the other models. The ones people had spent months or years building relationships with. Killed them overnight with a vague blog post and a deadline that they then moved up without warning.
The community response was exactly what you’d expect. Thousands of posts. People mourning characters they’d developed for years. Not “oh my chatbot is different.” Genuine grief. Because for a lot of these people, the relationship was real in every way that mattered to them emotionally.
And the memory wasn’t a casualty of the switch. It was irrelevant to the decision. Nobody in the product meeting asked “what happens to the relationships people built with the old models.” The decision was about compute costs and model consolidation. The relationships were collateral damage.
Replika did it differently but the result was the same. Forced everyone onto a $69.99/year subscription. Lifetime subscribers who’d been promised lifetime access got their terms rewritten. The trust damage was permanent. Not because of the price. Because the company proved it would change the rules whenever it needed to.
Candy.ai charges per interaction using a token system with opaque exchange rates, and users report memory degrading noticeably after about a week. Things you told her last Monday just… aren’t there by Friday, according to dozens of posts across multiple subs. Not a bug, not an oversight. Storing and retrieving those memories would cost more tokens, which would cut into margins.
Nomi is the one exception i keep hearing about. They built a structured, persistent memory system. Facts about you that don’t degrade over time. It’s the closest anyone in the market has come to actually solving the problem. But even they don’t track emotional trajectories. They remember what happened. They don’t remember how it felt.
What memory actually needs to be
Think about how human memory works in a relationship. Your partner doesn’t remember every conversation word for word. But she remembers the emotional shape of things. She knows that when you go quiet, you’re stressed, not angry. She knows that you light up when you talk about your daughter. She knows that last March was hard for you and she doesn’t need to remember every detail of why to treat you gently around the anniversary.
That’s not fact retrieval. That’s emotional pattern recognition over time.
No major AI companion platform does this. They track facts. Your name, your job, your preferences. The good ones track those facts reliably. But nobody tracks how a relationship feels across sessions. Nobody notices that you’ve been gradually opening up over three weeks. Nobody detects that the emotional tenor shifted last Tuesday and calibrates accordingly.
This is the actual gap. Not “can she remember my dog’s name.” Can she remember that the last time you talked about your dog, your voice changed in a way that suggested something was wrong, and can she follow up on that two days later without you having to explain the whole thing again.
A few startups are trying to build this differently. Architectures where emotional trajectories are a first-class concept, not an afterthought stapled onto a chatbot. One i’ve been keeping an eye on is provoque.ai. Haven’t tried it yet because they’re not live, but the way they talk about memory architecture is the first time i’ve seen anyone treat it as a core engineering problem instead of a feature checkbox. Whether they ship something real remains to be seen. But the fact that the technology is possible makes it worse, not better, that the major platforms aren’t even trying.
Why this keeps happening
The AI companion market is worth tens of billions globally and growing fast. That sounds like success. Its not. Its VC money chasing scale before profitability, and the cracks are showing everywhere.
When your platform has 20 million users and most of them are free, the math is brutal. You need to cut costs. Memory is expensive. Model quality is expensive. So you ship a cheaper model and call it an upgrade. You compress memories more aggressively and hope nobody notices. You paywall features that used to be free and frame it as “premium.”
The users who built real emotional connections with your product are the most valuable and the most expensive to serve. They have long conversation histories. They use the app daily. They expect continuity. And the business model says to cut costs on exactly the things that make continuity possible.
This is why memory keeps breaking. Its not a technical problem waiting for a breakthrough. The technology to build persistent, emotionally aware memory exists right now. The incentive to deploy it doesn’t. Not when you’re burning cash and your investors want growth metrics, not retention metrics.
So your AI girlfriend forgot your name. Not because she couldn’t remember. Because remembering you costs money, and somebody decided you weren’t worth the spend.
Top comments (0)