Memory is the product.
An assistant that forgets you between conversations cannot act for you. Memory that is kept well changes what it can be trusted to do.
Most chat assistants begin every conversation from nothing. You explain that you are vegetarian, that your father's follow-up is in November, that you fly out of Bengaluru, and the next day you explain it again. The model is capable. It simply has no past with you.
That limits more than convenience. Without memory, an assistant cannot act on your behalf, because acting well depends on knowing your context. Booking a table means knowing who is coming and what they eat. Paying a bill means knowing which account it comes from. Reminding you of a deadline is only useful if it knows the deadline exists.
Remembering is easy. Remembering well is not.
Storing everything a person says is trivial. The hard part is deciding what a fact means, how sure we are of it, and when it stops being true. We think about memory in four properties.
- Provenance. Every fact should carry where it came from. "You told me on 2 September" is different from "a document suggested it", and both differ from a guess. When the assistant uses a fact, it should be able to show its source.
- Correction. People change their minds and their circumstances. Memory has to be visible and editable, and forgetting has to be real: deleted means gone, not hidden.
- Scope. What you tell your personal assistant should not appear in a shared workspace, and company knowledge should follow company permissions. Memory needs boundaries as clear as file permissions.
- Expiry. Many facts have a shelf life. A flight lands, an address changes, an offer ends. A memory system that never lets facts age will eventually act on something stale, with confidence.
The test of memory is not recall. It is whether the assistant uses the right fact at the right moment, without being asked twice.
Memory should be used quietly
The best use of memory rarely announces itself. When you ask for a restaurant on Friday and the options already exclude mushrooms and keep Monday evening free, memory has done its job. When you ask "what was that hotel?" on the web and get the one you sent on WhatsApp on Tuesday night, it has done its job across surfaces.
We also think an assistant should be restrained about what it shows of its memory. Pointing to the source of a fact helps you trust it. Narrating everything it knows about you does not.
Memory as the base layer
This is why memory sits underneath everything in AskGogo, and why we are developing it as its own layer, SecondBrain. Reminders, briefings, documents, travel and browser tasks all read from it and write back to it. A task that finishes leaves the memory slightly better informed, so the next task needs fewer questions.
Over months, that compounding is the difference between a tool you use and an assistant you rely on.