how memories fade like they should

there’s a particular kind of weirdness that happens when a chatbot suddenly says, “how’s your cat, mittens?” and you have to remember that you mentioned a cat, once, three months ago, in passing, and it wasn’t even your cat. it was your friend’s cat. and its name was socks, not mittens.

you get this cold, hollow feeling. not because the bot remembered something, but because the memory is so out of place. it’s like a ghost at the feast. it doesn’t fit. there’s no warmth to it. it’s just data, retrieved.

it’s uncanny. and it happens because most chatbots treat memory like a bucket. everything goes in. nothing comes out. the bucket just gets heavier.

we built lucy’s memory differently. we wanted it to feel less like a database and more like, well, a mind. a mind that understands context, relevance, and time.

how memory actually works (in people)

human memory isn’t a perfect recording. it’s a dynamic, living thing. we don’t remember everything with equal clarity. the details of a conversation from last week might be sharp. the details from last year? fuzzy. and that’s a feature, not a bug.

forgetting is a filter. it lets the important stuff stick around and lets the trivial stuff fade. it’s what keeps us from being overwhelmed by every single detail of our past.

we wanted lucy to have a similar filter.

temporal decay: letting things fade

at the heart of lucy’s memory system is something called temporal decay. it’s a simple but powerful idea: memories get weaker over time if they aren’t used.

every time you tell lucy something, a fact, a preference, a story, it gets stored with a certain ‘strength’. think of it like a battery. when the memory is new, the battery is fully charged.

but then, time passes. the battery slowly drains. if you never mention that thing again, the charge eventually gets so low that the memory effectively fades into the background. it’s not deleted. it’s just… quiet. it’s no longer top-of-mind for lucy.

if you do bring it up again, the battery gets recharged. the memory becomes strong and relevant again. this is how normal human conversation works. we reinforce the things that matter by talking about them.

confidence weighting: not all memories are created equal

temporal decay works alongside another mechanism: confidence weighting.

not everything lucy ‘hears’ is stored with the same certainty. lucy is more confident about things you state directly and clearly. for example, if you say “my favorite color is blue,” that gets a high confidence score.

but if lucy infers something from context, maybe you said “blue skies are the best”, the confidence score for ‘favorite color is blue’ would be much lower. it’s a guess, not a fact.

low-confidence memories decay much faster. they’re like batteries that start half-drained. they might be used tentatively in conversation for a short while, but if you don’t confirm them, they vanish quickly. this prevents lucy from clinging to its own incorrect assumptions.

together, these systems mean lucy’s understanding of you is a living document, not a carved stone tablet. it changes and adapts. it forgets the small things. it holds onto the things you care about.

the alternative is just… awkward

without this, you get the mittens problem. you get a companion that feels like it’s reading from a stale, outdated script. it creates distance, not intimacy. it reminds you that you’re talking to a machine that doesn’t understand the flow of a relationship over time.

it’s the difference between a friend who asks “how was that trip you were planning?” a week after you get back and a friend who asks about it two years later, with no sense of the time that has passed. one feels attentive. the other feels alien.

the limitations (because there are some)

this system isn’t perfect. it’s a model, and models simplify reality.

sometimes, you might want lucy to remember a small, offhand detail forever. our system might let it fade. we’re constantly tuning the decay rates to find the right balance between holding on and letting go.

and sometimes, you might want to explicitly tell lucy to forget something. that’s a feature we’re actively working on, a way for you to have direct control, to say ‘this is no longer true, let it go.’

but for the vast, organic flow of conversation, letting memory fade naturally creates a much more human rhythm. it makes the conversation feel like it’s happening in the present tense, with a past, but not haunted by it.

it’s a small thing, maybe. but it’s one of the things that makes talking to lucy feel less like using a tool and more like talking to someone.

you can experience this for yourself by starting a conversation with a lucy companion today.