Why does AI forget between chats?
Because a model has no memory of its own. What looks like remembering is context, and context empties.
A language model has no memory; every conversation starts from nothing. What looks like remembering is a context window filled with the current chat, and when a new conversation begins, that window empties. Cloud assistants that seem to recall you write notes about you into a database and paste them back in; the model itself still learns nothing. Local tools keep history on your disk, so what is remembered stays on your machine.
Every chat starts from nothing.
The model you talk to is a finished file of numbers, and running it changes nothing in the file: inference reads and answers, it never learns. Everything the model knows about the current conversation is whatever fits in the context window at that moment. Close the chat, and there is nothing to carry over, because no part of the model was changed by the talk. This surprises people because the model reads so fluently within a session, but fluency inside the window and memory across windows are different facilities, and only the first one exists.
What remembering actually is.
Products that greet you by name last week's project are not running a model that remembers. They store notes about you in a database, select some, and put them at the front of the context before your first message. The model performs the same unchanging weights in both cases; the continuity is a feature of the storage around it. That distinction matters when memory goes wrong: a note written once keeps resurfacing, and the model cannot tell you which of its "memories" came from where, because they are all just text it was handed.
Where local memory lives.
A local tool can offer the same continuity, and the honest way to think about it is custody. History kept in files on your disk behaves like any other file: it backs up when you back up, moves when you move it, and disappears when you delete it. Re-injecting old notes also consumes context, so a well-built local tool treats memory as a budget like any other. The model still starts fresh every time; what persists is yours, in a format you can open and a place you control.
The limits of the notes approach.
Injected memory has failure modes worth knowing, because they explain odd behavior in assistants that seem to remember. A note written once keeps resurfacing in unrelated conversations, since nothing marks it expired. Selection is invisible: you cannot see which notes were pasted in, so you cannot correct the record the model is working from. And notes consume the same context budget that documents and questions share, so a crowded memory silently crowds out the material you actually asked about. A tool that treats memory as editable text you can inspect, rather than a hidden store, sidesteps most of this by making the notes ordinary files.
What continuity is worth wanting.
The useful version of memory is narrower than the fantasy. Projects that persist: notes, sources, and past answers kept on disk and re-opened when you return. Preferences that stick: a citation style, a level of detail, kept as settings rather than learned behavior. Context that carries within a session, which the context window already does. The model learning from you remains off the table in every case, local or cloud, and that is mostly fine: for research work, a library on disk with a model that reads it well beats a model that vaguely recalls you. Local-first design is largely this arrangement made deliberate.