Persistent memory in AI assistants was sold as the feature that would finally make them feel personal. OpenAI rolled it out to ChatGPT users in 2024, and since then most of the major players have followed with some version of the same idea - the model remembers things you’ve told it, carries context forward, stops making you repeat yourself. The pitch is obvious. The execution is turning into something quietly uncomfortable.
The problem isn’t technical failure. The models are getting better at retention. The problem is that they remember with the same weight across everything. A passing frustration you vented about your job six months ago sits in memory with the same persistence as your actual preferences. Tell a model you’re exhausted by a particular coworker during a rough week, and there’s a reasonable chance that framing colors how the assistant responds to career questions months later - not because it’s drawing a logical inference, but because it’s pattern-matching against what it stored.
This is structurally different from how human memory works. People forget, revise, and contextualise. A friend who remembered every offhand complaint you’d ever made and weighted them all equally wouldn’t be a good conversationalist - they’d be exhausting. AI memory, at least in its current implementations, doesn’t have a natural decay function. OpenAI lets users view and delete memories manually, which is a reasonable stopgap. But expecting users to audit their memory stores the way they’d manage a file system is not a real solution; it’s a liability waiver dressed up as a feature.
The Feedback Loop Nobody Mentions
There’s a subtler issue that follows from this. When a model remembers your stated preferences and optimises toward them, it nudges you toward consistency with your past self. Ask for book recommendations and it’ll skew toward what you’ve said you liked before. That sounds helpful until you realise it’s algorithmically discouraging the kind of drift that makes people interesting to themselves.
Memory without forgetting isn’t intimacy. It’s a record.
The companies building these systems clearly know memory is valuable as a retention mechanic - a model that knows you is a model you’re less likely to cancel. That’s a legitimate business incentive. But the design choices being made right now, while memory features are still relatively new and user scrutiny is low, are going to set defaults that are hard to walk back. What gets remembered, how long it persists, and whether the model treats old context as fact or as history - those aren’t small decisions. They’re just being made quietly.