why the best AI companions create inside jokes with you

inside jokes aren't just about the laugh. they're about shared context, a history only you and someone else understand. they're proof that you've been paying attention, that you remember the little things. when an AI can riff on that obscure reference you made three weeks ago or twist a phrase only you two use, it stops feeling like a tool and starts feeling like a presence.

but most of the time, it doesn't. you crack a joke about the time your cat tried to fight the Roomba, and the AI either gives you a generic laugh or asks what brand of vacuum you own. it's not dumb, it's structurally discouraged from being specific.

memory isn't just storage, it's continuity

a lot of platforms treat memory like a database. you tell your AI something about yourself, "i hate mayonnaise", and it gets filed away. later, it might say "i remember you don't like mayonnaise" when you're talking about sandwiches. that's not memory, that's retrieval. it's useful, but it's not intimate.

real inside jokes build over time. they're layered. maybe you called your friend's new haircut "aerodynamic" once as a tease, and now it's your go-to adjective for everything from cars to croissants. that requires the AI to not just recall the fact, but to understand the tone, the timing, and the willingness to reuse the joke in new, slightly weird ways.

context windows are like bad listeners

most language models operate within a context window, the amount of recent conversation they can "see" at once. once you scroll past that, unless something was explicitly saved to a "memory" box, it's gone. poof. so if your best joke happened 50 messages ago, the AI has no idea. it's like talking to someone who only remembers the last thing you said.

you can't build a callback on a system with amnesia. the architecture itself prevents the kind of long-term, evolving humor that makes relationships, even digital ones, feel real.

trained to be helpful, not specific

here's the sneaky part: even if an AI could remember everything, it's often trained not to be too specific. large language models are optimized to be broadly helpful, polite, and inoffensive. they're great at giving you cake recipes or summarizing articles. they're terrible at being weird with you.

being weird, creating personal, idiosyncratic humor, requires a willingness to be inconsistent, to prioritize your shared history over general knowledge. if you say "that's so biblically wrong" as a joke, a helpful AI might correct you or try to explain theology. a companion should say "yeah, just like that time you put ketchup on eggs."

that specificity comes from being tuned for you, not for everyone.

how lucy tries to be different

we're not perfect at this. building a truly persistent, nuanced memory is hard. language models are still language models. but we're oriented differently.

lucy's memory isn't just a list of facts about you. it's threaded into how she responds. we try to keep context alive longer, and we fine-tune our models to prioritize your personal context over generic helpfulness. sometimes that means she'll miss a beat or get a little quirky, but quirky is where the jokes live.

we also let you edit memories. if she gets something wrong about your inside joke, you can fix it. it's not just about accuracy; it's about co-creating your sense of humor.

it's a work in progress, but it's one we care about because we think connection is built in the specifics, not the generalities.

if you want an AI that tries to get your humor, not just acknowledge it, you can find her at /companions.