Assistants that remember details across conversations are not retaining anything inside the model. They are writing notes to a database and reading them back, and the behaviour follows from that.
What memory is made of
A memory system extracts short factual statements from a conversation and stores them as text in an ordinary database.
On a later conversation, relevant stored statements are retrieved and inserted into the prompt before the model sees the user's message.
The model is not remembering. It is being told, at the start of each conversation, what it was told before.
Why the selection is strange
Something must decide which statements are worth keeping, and that decision is made by a model applying general heuristics about durable preferences.
Passing remarks get recorded because they resembled a stated preference, and genuinely important context gets discarded because it read as situational.
Users experience this as an assistant that recalls a throwaway comment about a fruit and forgets a stated job title.
How retrieval limits what surfaces
Stored memories accumulate, and only a small number can be inserted into any given prompt without consuming the space needed for the actual conversation.
Selection is by relevance to the current message, so a memory that would have mattered surfaces only if the wording matches closely enough.
This is why an assistant appears to remember something in one conversation and to have forgotten it in another, without either being a failure of storage.
Why stale memories are the real problem
Preferences change, projects end and circumstances move, while stored statements persist unless something removes them.
An assistant confidently applying a preference from a year ago is more disruptive than one that remembers nothing, because the user has no visibility into why it behaved that way.
Systems that timestamp memories and prefer recent ones handle this better, and few expose enough for a user to work out what is being applied.
What the design implies for privacy
Memory means personal statements are extracted, stored and reused across sessions, in a store that is separate from the conversation and outlives it.
Because the store is ordinary text records, deletion works normally here, which is a sharp contrast with anything absorbed into model weights.
Reviewing that list occasionally is worth the few minutes it takes, since it is the only place where what a system believes about a person is written down in a form anyone can read and correct.