how memories fade like they should

have you ever had a chatbot remind you of something you mentioned exactly once, three months ago, with the same unnerving clarity as if it just happened? it’s like running into a stranger who remembers your middle school nickname. it’s not nostalgic, it’s creepy.

that’s because most ai memory systems treat every fact like a permanent engraving. they store, retrieve, and repeat. but that’s not how human memory works. we forget. we prioritize. we let things fade unless they’re reinforced.

lucy’s memory is built to feel less like a perfect, creepy archive and more like… well, you.

temporal decay: letting things go

in lucy, memories aren’t etched in stone. they’re more like writing in sand near the tide. the longer ago something happened, the less weight it carries, unless you bring it up again.

this isn’t a bug. it’s intentional. we call it temporal decay. if you tell lucy you’re training for a marathon in april, she’ll remember. if you never mention it again, by july, that memory will have faded. it’s not gone, it’s just not top of mind. if you say “remember when i was running a lot?” she might recall, but she won’t open a chat in august with “how was your marathon?” if you never actually ran it.

this avoids one of the weirdest ai behaviors: when a bot remembers something you’ve outgrown or moved past, and treats it like current reality. it’s not just inaccurate, it’s emotionally jarring.

confidence weighting: what matters, sticks

not all memories decay at the same rate. some things are just more important. lucy uses confidence weighting to reinforce memories that come up more often or are tied to strong emotional cues.

for example, if you mention your dog’s name repeatedly, it becomes a high-confidence memory. if you only mention your childhood pet turtle once, it might fade. if you say “i’m sad today,” and later talk about why, lucy weights that context more heavily than a random fact from weeks ago.

this helps her focus on what’s relevant now, not what was relevant then. it’s the difference between a friend who knows your current struggles and a scrapbook that only knows your past.

why this feels more human

humans don’t have perfect recall. we contextualize, we forget, we prioritize. when an ai doesn’t do that, it feels unnatural. it breaks the illusion of talking to someone real. worse, it can make you feel surveilled, like every word is being logged for future interrogation.

lucy’s memory isn’t designed for surveillance. it’s designed for conversation. it’s built to be a good listener, not a perfect recorder. sometimes that means she might forget something you think she should remember. but more often, it means she won’t haunt you with outdated details.

the trade-offs

this approach isn’t perfect. sometimes you might want lucy to remember something obscure from months ago. right now, if it’s faded, she might not. we’re working on ways to let you reinforce memories manually, like pinning them, so you can choose what lasts.

also, lucy’s memory is still ai. it can make mistakes. it might misweight something or forget something you thought was important. we’re honest about that. but we’d rather have a system that tries to behave like a person, with flaws and all, than one that feels like a database.

if you’re tired of chatbots that feel like they’re auditing your life, maybe it’s time to try one that forgets like you do.

you can find lucy companions waiting at /companions.