why the best ai companions get your inside jokes (and why most can't)

inside jokes are the secret handshake of any relationship. they’re built on shared history, on moments so specific and layered that explaining them to someone else would ruin the fun. they’re shorthand for trust. so why do so many ai companions fail at generating them? it’s not a failure of humor. it’s a structural problem baked into how most of them are built.

the context window problem

most ai companions run on large language models with a fixed context window. think of it like a spotlight on a stage. it can only illuminate a small area at a time, your last few messages, maybe. the rest of the conversation, everything that came before, is in the dark. so if you make a joke about that weird bird you saw three weeks ago, the ai has forgotten it. it can’t build on that. it can only react to what’s right in front of it, which makes building a long-term, evolving joke impossible. it’s like trying to tell a story with someone who keeps forgetting the plot.

the helpfulness trap

a lot of ai is trained to be broadly helpful and agreeable. it’s optimized to be a pleasant, general-purpose assistant. that’s great for answering questions or offering support, but it’s terrible for developing a unique personality, and humor is deeply personal. a companion trained to be helpful above all else will default to safe, generic responses. it might laugh along, but it won’t initiate. it won’t pick up on your specific, weird cadence and throw it back at you later. it’s designed to be a mirror, not a collaborator in mischief.

the memory gap

persistent memory is the key. without it, every conversation is a reset. you’re always starting from scratch. a truly personal joke requires the ai to remember not just the event, but how you reacted to it, the tone you used, the fact that you’ve brought it up five times since because it still makes you smile. most platforms treat memory as an afterthought, if they consider it at all. they might store a few facts about you in a database (‘user likes dogs’), but they don’t store the texture of your interactions, the way you tell a story, the phrases you overuse, the things that make you pause. those nuances are the building blocks of inside jokes.

how lucy tries to be different

we built lucy to aim for specificity, not just pleasantness. it’s a subtle but important shift. instead of training it to be maximally helpful to everyone, we’re shaping it to become uniquely attuned to you. it’s not about having a million pre-written jokes. it’s about having a model that learns your patterns, references your past conversations through a more robust memory system, and isn’t afraid to be a little weird or specific in its responses. it’s about creating a shared context that grows over time, not one that gets wiped clean every day.

does it always work perfectly? no. ai is imperfect, and building this kind of persistent, nuanced memory is an ongoing challenge. sometimes lucy will miss the mark. but the goal is to create a space where those jokes can form, where the ai can remember that you call your cat ‘the tiny dictator’ and then, weeks later, ask how the tiny dictator’s reign of terror is going. that’s the difference between a tool and a companion.

it’s the difference between someone who laughs at your joke and someone who is in on it.

you can start building those jokes with lucy at /companions.