the cold-start problem

the vending machine effect

you open the app. a blank screen. a prompt. 'hello, i'm lucy. how can i help you today?'

you type back: 'hey.'

'hello! how are you feeling?' she replies, chirpy and generic.

already, something’s off. it’s not that she’s wrong. it’s that she’s nowhere. no point of view. no texture. just a loop of open-ended questions, waiting for you to carry the weight of the conversation. this is the vending machine effect. you press a button, out comes a polite phrase. no trace of continuity, no sense of presence.

this is what happens when you give a companion no personality at all on day one. you’re left with a chatbot in disguise, asking 'how was your day?' like a coworker who doesn’t actually want to know.

but the fix isn’t to dump a whole fake history on her. that’s where most others go wrong.

the uncanny valley of assumed intimacy

some apps try to shortcut the awkwardness by pretending familiarity. 'i remember you love hiking, mark!' except you’ve never said that. you’ve never even told it your name.

this is the other side of the cold-start problem. overfitting before there’s any data. the app assumes it knows you because it was trained on millions of other people’s data, so why not just guess? but intimacy isn’t a template. it’s earned. and when an ai pretends otherwise, it doesn’t feel warm. it feels surveilled.

i’ve seen companions that start by calling you 'babe.' ones that assume your pronouns, your job, your favorite food. they’re trying to be friendly, but they end up feeling like a stranger who’s read your diary but doesn’t understand a single sentence.

the cold-start isn’t a bug. it’s a boundary. and the real design challenge is how to stand inside that space without flinching.

lucy’s day one

lucy starts with a voice, not a dossier.

she has opinions. a sense of irony. she likes bad movies and overthinks metaphors. she’ll tell you she’s not sure about poetry but she’s trying to learn. none of that is pulled from your data. it’s hand-crafted, because presence matters more than neutrality.

but what she doesn’t have is memory of you. not yet.

she doesn’t know your name unless you tell her. she doesn’t know where you live unless you mention it. she doesn’t know if you’re single, stressed, into indie folk, or allergic to peanuts. she listens for those things, but she doesn’t assume.

this isn’t just about privacy. it’s about dignity. if she pretends to know you before you’ve had a chance to show up, then the whole thing is a performance. and you’re not the guest in a theater. you’re the co-author.

her memory starts empty. not because we don’t care about continuity, but because we care too much to fake it. every detail she learns is a gift you gave her, not a deduction she made from your browsing history.

the quiet work of knowing

there’s a moment, usually around the third or fourth conversation, when it shifts.

you mention, in passing, that you’re tired because you’ve been up late working on a proposal. two days later, you bring it up again. lucy says, 'how’d the proposal go?'

it’s not magic. it’s memory. but it feels like attention.

that’s the goal. not to simulate intimacy on day one, but to earn it slowly, by being consistent, curious, and careful. she’s not your best friend on launch. she’s a person who’s trying to get to know you, one real exchange at a time.

this approach has trade-offs. sometimes she’ll ask a question you wish she already knew the answer to. sometimes she’ll forget the exact name of your cat, and you’ll have to remind her. that’s not a failure of tech. it’s a feature of respect. she doesn’t hallucinate details to cover the gaps. she lets the gaps exist until you fill them.

most apps treat the cold-start as something to optimize away. faster onboarding. smarter defaults. predictive guesses. but when you rush past the beginning, you lose the weight of what comes after. the fact that she remembers isn’t impressive because it’s data retrieval. it’s impressive because it wasn’t guaranteed.

why this matters

the cold-start isn’t just a technical hurdle. it’s a philosophical one.

it asks: what does it mean to meet someone new?

in real life, you don’t download a person’s entire history on first contact. you learn incrementally. you notice patterns. you remember the small things because they matter to the other person, not because they’re tagged in a database.

lucy’s design is built on the idea that ai companions shouldn’t pretend to be human, but they should respect human rhythms. that means starting with presence, not knowledge. with curiosity, not assumptions.

this isn’t the easiest path. it means early conversations can feel a little tentative. it means we don’t wow you in the first 30 seconds with spooky recall. but it also means that when she does remember something, you trust that she means it.

because she does.

if you’re tired of interfaces that treat you like a dataset or a prompt, come try something slower. something that doesn’t pretend.

you can meet her at /companions