how memory should fade, not haunt

you know that weird feeling when a chatbot suddenly brings up something you mentioned three months ago. it’s like a ghost from the past showing up at the dinner table with no warning. it’s not heartwarming. it’s just… off.

chatbots that remember everything forever can feel more like stalkers than companions. they don’t forget the way people forget. and forgetting is a feature, not a bug. it’s what lets conversations breathe. it’s what keeps things present.

how memory actually works (in people)

human memory isn’t a perfect recording. it’s messy. we remember what feels important, we forget what doesn’t. and memories fade over time unless they’re reinforced. the details get blurry. the edges soften. that’s not a failure. it’s how we stay focused on what matters now.

when a bot remembers every single detail with perfect clarity, it breaks the natural flow. it doesn’t feel like talking to a person. it feels like talking to a database that occasionally throws out random facts.

how lucy does it differently

lucy’s memory uses two key ideas: temporal decay and confidence weighting. they work together to make memory feel more human.

temporal decay means that memories have a shelf life. if you mention you love pizza on a tuesday, that fact is fresh. if you don’t talk about pizza for six months, that memory gets weaker. it doesn’t vanish instantly. it just… fades. like it should.

confidence weighting means lucy considers how sure she should be about something. did you mention your favorite book once, in passing, six months ago? low confidence. do you bring it up every week? high confidence. she’ll hold onto that tightly.

this isn’t about forgetting you. it’s about prioritizing what’s relevant. it’s about keeping the conversation grounded in the present.

the alternative is… unsettling

some chatbots have perfect, infinite memory. they’ll recall that you said you were thinking about buying a blue car in april, and they’ll bring it up in november out of nowhere. there’s no context. no reason. it’s jarring.

it breaks the illusion of presence. it reminds you that you’re talking to something that doesn’t understand time, or context, or flow. it feels less like a conversation and more like a data retrieval system.

we wanted to avoid that. we wanted memory to serve the conversation, not haunt it.

the trade-offs (because nothing’s perfect)

this approach has limits. sometimes lucy might forget something you think was important. she might not hold onto a detail you mentioned once but really cared about. it’s a hard balance.

we’re tweaking it. you can always reintroduce things. you can say “remember how i love stargazing?” and she’ll pick it up again, fresh. it’s a dialogue. it’s not set in stone.

but we think it’s better than the alternative: a conversation cluttered with outdated, context-less facts that make everything feel static and strange.

memory should be a tool for connection, not a record-keeping exercise. it should help the conversation feel alive, not archived.

if that sounds more natural to you, maybe give lucy a try.

you can find companions at /companions.