How character memory is supposed to feel
If you live in long character chats, you already know the failure mode. You mention something that matters on day one. Forty messages later, the character acts like you never said it. Or the opposite happens: they cling to a throwaway joke like it’s a life fact.
We’re building conversation memory so chats keep the useful stuff without treating every “lol” like a biography entry. This post is about how that should feel when you’re using it. Not the machinery under the hood.
Save, then bring back
Memory has two jobs. First, notice things worth keeping. Preferences. Relationship vibe. Story details that should still matter later. Second, bring a small, relevant slice of that back when it helps the next reply.
Temporary things can live for a bit and drop off when they’re done. Waiting on a file. A trip this week. A mood that was only about today. Durable things stick around: what you like to be called, a running joke that still lands, the plot beat you both care about.
The goal isn’t a perfect archive. It’s a character who doesn’t blank on the parts that make the chat feel like yours.
You should be able to see it
If something gets saved, you shouldn’t have to guess. You’ll be able to open the chat’s learned memories, delete anything that feels wrong, clear the list, or turn learning off for that conversation. That control matters more than a clever black box.
People trust memory when they can correct it. They bounce when the chat “remembers” something creepy or wrong and there’s no way to fix it.
Stays in this chat
Memory is scoped to that conversation. What you tell Character A does not bleed into Character B. If you want standing instructions everywhere, that’s still the kind of thing you set on purpose in settings. Auto memory is what the chat learns on its own.
That boundary is part of the product feel. Different chats can be different vibes without getting tangled.
What we’re not trying to do
We’re not trying to turn every conversation into a dossier. We’re not trying to surprise you with “I remember everything.” We’re trying to make long chats less brittle: fewer blank stares, fewer weird stuck facts, more continuity when it counts.
This is still baking. We’re tightening what gets saved and how it comes back into replies. When you can try more of it in the product, we’ll say so. Until then, the north star is simple. Keep what matters. Skip the noise. Let you see and edit the result.
A quick example of the feel
You mention you hate cilantro. Later you’re talking about dinner plans and the character steers around it without making a speech about how great their memory is. That’s the win. Quiet continuity beats a brag.
Bad memory feels like either amnesia or a scrapbook that won’t shut up. We’re aiming for the middle: useful, light, correctable.
If you’re chatting today
Keep talking the way you already do. Mention the durable stuff when it matters. Don’t perform for the system. When the editable memory surface is in your hands, tidy anything that feels off and keep going.