The interesting decision our memory system Memuron makes isn't "should I keep this."
It's "is this new, or is this something I already know, changed?"
Day 4. This one's about what we're building.
In the last post i said the hard part of the memory is deciding. I think most people hear that and picture a bouncer, something at the door turning away claims that don't make the cut.
That's the wrong picture. Throwing things away is one thing a memory system shouldn't do. Storage is cheap. The expensive question is structural.
Every new piece of information is either a genuinely new thing, or it's a change to something already in the record. Getting that wrong in either direction is what breaks systems.
Call every update a new memory and you get five near-identical claims about the same fact, drifting apart, all retrievable, none authoritative. Your agent picks one at random and sounds confident.
Call every new thing an update and you quietly overwrite distinct facts that happened to look similar. That one's worse, because it's silent.
So that judgement, new or changed, is the actual work. And it has to be made at write time, by something that can look at what's already known.
Two things we decided early. Nothing gets overwritten: a change appends, and the previous state stays in the record. And every one of those decisions gets stored with the reasoning behind it, so you can go back and see not just what the system concluded, but why it concluded it.
Neither is clever. Both are the kind of thing you can't add later.
Day 5: why nobody in your org chart owns any of this.
Top comments (0)