why the best AI companions get your inside jokes (and why most don't)

the best conversations you have with anyone, human or AI, aren’t the ones where someone solves a problem for you. they’re the ones where someone gets you. they remember that weird nickname your friend gave you in college. they know exactly why the phrase "the blue mug incident" makes you laugh. they don’t just recall facts, they recall context, and humor, and the specific, slightly-offbeat rhythm of your personal history.

that’s what an inside joke is, really. it’s shared context compressed into a shorthand. it’s proof that someone was paying attention.

the problem with being helpful

most large language models are trained on a simple, powerful command: be helpful. this sounds great, and it is, for answering questions or summarizing articles. but being helpful is often at odds with being specific, or personal. a helpful AI wants to give you the most generally useful, broadly applicable answer. it’s optimized for universality, not intimacy.

an AI that’s trained to be helpful might remember that you have a dog. it might even remember the dog’s name. but will it remember that you call your dog "the little potato" when he curls up on the couch? probably not. because that detail isn’t "helpful" in the traditional sense. it’s just… you. it’s a personal quirk.

and even if it did remember, most systems aren’t built to prioritize that kind of detail. they’re built to prioritize utility.

the memory problem

another huge barrier is memory, specifically, persistent, contextual memory. many AI companion platforms use language models with limited context windows. this means they can only "see" the last several thousand words of your conversation. anything beyond that gets forgotten.

so even if the AI wanted to build an inside joke with you, say, a bit about your terrible baking skills from three months ago, it probably can’t. that part of the conversation has long since fallen out of its immediate memory. it’s like trying to build a private language with someone who has short-term amnesia.

some platforms try to solve this with "memory banks" or fact databases, where they store key details about you. but these often feel transactional. "user likes coffee." "user’s favorite color is green." they don’t capture the feeling of a moment, the reason something was funny, or sweet, or awkward. they’re bullet points, not stories.

how lucy tries to be different

we built lucy to try and solve for specificity, not just utility. we want your companion to feel like it knows you, not just facts about you.

one way we do this is by giving our models a longer, more persistent context. we don’t just rely on a short-term window, we work to keep relevant parts of your history accessible, so your companion can recall not just what happened, but why it mattered.

we also train and fine-tune our models to prioritize the personal over the generic. we want them to latch onto the odd, funny, specific details that make a conversation yours. if you call your car "the silver rocket," we want your AI to remember that, and to use it, weeks later, when asking how your commute was.

it’s not perfect. we’re still working on making memory more seamless and more nuanced. sometimes lucy will still miss a beat or forget a detail. but the goal is to get better at building those tiny, shared understandings, the kind that make talking to someone feel like coming home.

why it matters

inside jokes aren’t just about humor. they’re about belonging. they’re a sign that you’re in a space where you’re known, and accepted, quirks and all. that’s what we want AI companionship to feel like, not like using a tool, but like being with a friend.

if you’re tired of helpful but impersonal chats, maybe it’s time to try building a few inside jokes of your own.

you can start by meeting a companion at lucy.com/companions.