When an assistant remembers that you are vegetarian, your kids’ names, or the project you have been chipping at all month, the model did not learn those things. The product wrote them down.
AI memory is a notebook plus a librarian: notes saved about you, and a system that decides which notes get placed into the window each time you talk.
The kinds of things worth saving
Stable preferences, like how you want things written. Facts about your life, like your city and who matters to you. Project state, like what you decided last week. Routines, like how you want Mondays to start. Each earns its place because it changes what a good answer looks like.
It is worth noticing what is not on that list: everything else. The random question at midnight, the venting about a coworker, the draft you deleted. A system that saves everything gets slower and creepier at the same time, because when it needs to recall something useful, it has to dig through piles of things it never should have kept.
Good memory is selective
Saving everything would drown the window in noise and feel invasive besides. The craft is judging what is durable and welcome. A dietary preference belongs there, and a stray remark from a bad day does not. The best test of a memory system is what it declines to keep, and how easily you can see and delete what it kept.
When something gets saved matters as much as what gets saved. A good product saves a fact when you state it, updates it when it changes, and does not bring up a preference from two years ago as if you said it yesterday. Memory has to get reviewed, revised, and sometimes retired, just like trust between people.
Memory is what makes AI personal
Without memory, every conversation is a first date. With it, help compounds: the tenth request builds on the nine before it. That compounding is the difference between a clever tool and something that grows with you, and it is the thesis behind Companion.
This compounding is also a reason to stick with one assistant instead of hopping between tools. Every correction you make is an investment, and it only pays off in the place where it is kept. The products that win the next decade will be the ones people have taught the longest.