Persistent memory in AI assistants has been quietly rolled out across most major platforms over the past year - ChatGPT, Claude, Gemini, and various wrappers built on top of them all have some version of it now. The pitch is simple: the model remembers facts about you across conversations, so you don’t have to repeat yourself. Reasonable enough. The problem is what the systems have decided is worth remembering.
The Profile They’re Building Isn’t Quite You
Most memory implementations work by extracting facts from your conversations and storing them as discrete notes - something like “user prefers concise responses” or “user is learning Spanish” or “user works in product management.” These feel useful in isolation. Strung together, they build a portrait that’s frozen at the moment of capture. You told the model something six months ago and it’s still anchoring to that, with no mechanism to notice that the context has drifted.
This is a fundamentally static model of a person applied to a dynamic situation. Preferences shift. Jobs change. The “product manager learning Spanish” might now be a consultant in a different industry who abandoned the language classes. The memory doesn’t expire, and in most implementations, users have to manually curate it - which almost nobody does consistently.
What It’s Actually Optimising For
The real beneficiary of persistent memory isn’t user experience - it’s retention. A model that seems to know you is harder to abandon than one you have to re-explain yourself to every session. That’s a legitimate product goal, but it’s worth being clear that it’s a product goal, not a usability breakthrough.
There’s also an asymmetry problem: the model accumulates information about you, but you have limited visibility into what it’s actually stored or how heavily it’s weighting specific facts. Some platforms expose a memory list; few make it obvious when a stored belief is actively shaping a response.
The Short Version
Memory without decay isn’t really memory - it’s a record. Human memory is reconstructive and context-sensitive; these systems are neither.
Where This Goes
The more sophisticated path - models that weight recent interactions more heavily, that flag when stored context seems to contradict new information, that treat memory as probabilistic rather than factual - exists in research but hasn’t made it into consumer products in any meaningful way. Whether that’s a hard engineering problem or just not a priority yet is an open question. The current implementations suggest the latter, but it’s not obvious what the incentive to change would actually be.