What I tried
Sam lets people share notes and tasks with each other. On the same day, two things went wrong. A meeting summary for a one-on-one with a coworker pulled in someone’s personal tasks. And a shared task, written in the first person about a family errand, convinced Sam that the person who shared it was part of that family.
What I found
Sam was reading shared words as if the reader had written them. To a model, “me” is whoever is talking, and nothing told it who was talking. The note was understood perfectly; the author was guessed wrong.
What I changed
Everywhere someone else’s words reach the model, they now arrive labelled with whose they are. Each memory Sam keeps records who it’s about and where it came from. And there’s now a set of tests built from shared notes, so this stays fixed.
What I’d do next
Give Sam a real sense of relationships, so it knows not just who wrote something but how that person relates to you.