A few things we figured out about long-term memory
Imagine running into an old friend you haven't seen in three years. They remember what you talked about last time. They remember where you were on the thing you were working on then. You can pick up where you left off. You get to reunite — not start over.
That's the most direct version we know of how memory makes a relationship. The difference between a stranger and an old friend isn't smarter and dumber. It's that one of them remembers you.
The AI you use today — most of it — is still a stranger to you, even after ten thousand uses. Every time you open it, it introduces itself. Every conversation, you re-explain. Every session ends, and the slate is wiped. It's not that it can't remember. It's that it was never designed to.
When we built Actopus, we got clear about three things on memory.
First. Memory isn't dumping everything into one pool. It's knowing what to keep and what to let go. Your preferences, keep. The way you like to work, keep. The progress you're making on what you're doing, keep. The wrong turn you took once — also keep. Not to blame you. So that next time, you don't take it again.
Second. Memory has structure. Short-term is a sticky note. Mid-term is your handbook. Long-term is a more and more complete picture of you. They each exist on their own; they reference each other; they don't get wiped by a single conversation; they don't disappear when you restart.
Third — and most important. The memory is yours, not its. You can see what it remembered. You can correct what it got wrong. You can make it forget. This is your partner, not your file cabinet.
The first time, it knows nothing about you. After a week, it starts to know your taste. After a month, it can carry more for you. After a year, we hope, it'll be one of the people you work with most.
What we're building isn't a smarter model. It's the part of memory that actually belongs to you.
—— Actopus Team