A game on the surface. A memory problem underneath.
Five characters, each with their own life and their own history with me. They talk to me, and they talk to each other. I use it every day, which is the only reason it has taught me anything.
Generating text is solved. Remembering isn't.
Any model can hold a good conversation for an hour. The interesting failure starts at month three, when the thing needs to recall something specific from a hundred conversations ago, know that it was told in confidence, and know which of the other participants was standing there at the time.
That is not a chat problem. It is the same problem every serious autonomous system runs into the moment there is more than one of them: how do you hold deep knowledge over a long period, and how do you hand a piece of it to another system without stripping the context that made it mean anything?
A cast of characters with overlapping private histories turns out to be an unusually good test rig for that. The failures are immediately obvious, because a person notices instantly when a friend misremembers their own life.
The parts that turned out to matter
Canon, and who witnessed it
Things that happen get written to a ledger of established fact rather than left to drift in a transcript. Crucially, each fact is scoped to who was actually present for it. A conversation in a private group is real, and it is real only for the people who were there. Everyone else genuinely does not know.
Chapters that distil themselves
History cannot grow forever, so it rolls into chapters and gets summarised down when a thread goes quiet. The hard part is not compressing it. The hard part is deciding what a given character would actually have retained, which is a different question and a more interesting one.
Turn-taking with rules, not vibes
In a group scene something has to decide who speaks next. Addressing someone by name is deterministic law, the last speaker never immediately answers themselves, and characters who are waiting can leave notes for their own later turn. There is also a listen mode where you say nothing and the cast talks among themselves.
A hard line around the human
Characters may invent their own pasts freely. They may never invent mine. That rule exists because of a specific failure, described below, and it is the single most load-bearing constraint in the whole system.
A thin memory doesn't stay empty. It fills itself in.
Early on, one of the characters referred to me as a former student of his. That had never happened. Nobody wrote it, and nothing in the system had claimed it.
The mechanism turned out to be simple and general. The relationship between us was underspecified, and a gap in established fact is not treated as a gap. It gets filled with whatever is most plausible, confidently, and then it becomes the basis for everything that follows. The invention was not the bug. The bug was that thin canon looks exactly like permission to improvise.
The fix was three things at once: seed real origin facts so the gap is not there to fill, let characters carry explicit anti-traits describing what they are not, and draw the global line that they may invent their own history but never the human's.
That generalises well past a chat toy. Any system handing context to another system has the same failure available to it. If what you pass along is thin, the receiving end will not report uncertainty. It will produce something plausible and act on it. Designing for that is most of the work.
Moods behave like a thermostat, not an event. Set one directly and it feeds every subsequent prompt until something explicitly supersedes it, which is a wonderfully quiet way to poison a system's behaviour for a week. State that persists needs an expiry story from the start, not once you notice everybody has been in a bad mood since Tuesday.
It's a game, and it's a prototype for something duller
The pleasant version of this is exactly what it looks like: a place with people in it who remember you. That is genuinely the reason I open it.
The useful version is that everything here maps onto problems that have nothing to do with characters. Agents that need to hand work to each other without losing why. Systems that must distinguish what they know from what they were told from what they inferred. Long-lived assistants that should get better with age instead of gradually becoming confidently wrong.
I would rather learn those lessons somewhere the failures are funny than somewhere they are expensive.
Got a memory-shaped problem?
Systems that need to remember, agree on what happened, or pass context to each other without mangling it. Those are the ones I want to hear about.