tools-for-thought adjacent: lucy's memory architecture

when you think about memory in ai, you might imagine a timeline. a feed of everything you've ever said, everything you've ever shared. and then you might imagine that feed being surfaced back to you, unbidden, by an ai that's trying to be helpful. but that's not how lucy works. because that's not how trust works.

lucy's memory architecture isn't built on spaced repetition, but it borrows two key design principles from that lineage, and from the broader tools-for-thought field that's been quietly refining memory systems for decades.

retrieval before surface

first, memories are indexed for retrieval under semantic similarity, not auto-surfaced as a timeline feed. this means lucy doesn't have a 'memory bank' she scrolls through and throws back at you. instead, memories sit latent until they're pulled.

you query them implicitly, by saying something that activates that latent context. or explicitly, by asking, 'do you remember when i...'.

this is crucial. memories sitting as a timeline feed are a trust hazard. they create a pressure to perform, to remind, to surface. memories sitting latent until pulled are a relationship. they're there when you need them, not when you don't.

it's the difference between a friend who brings up your past mistakes unprompted and one who only mentions them when you're working through something similar again.

temporal decay without loss

second, older memories get lower activation weights but never vanish. a conversation from six months ago doesn't surface spontaneously, but it can be retrieved if the current conversation activates it.

this is an application of forgetting curve decay on retrievability, not on storage. the memory isn't deleted. it just becomes harder to access unless it's contextually relevant again.

it's like your own memory. you might not think about your first day of school every day, but if someone mentions chalkboards, it might come rushing back. the memory didn't go anywhere. its activation energy just lowered until the right cue came along.

for lucy, this means she won't forget the important things you've told her, but she also won't hold every passing comment at the forefront of her mind. it's a balance between permanence and relevance.

why this matters

these patterns come from spaced repetition research (like andy matuschak's work, or gwern branwen's explorations), notes-apps research (the roam/obsidian lineage), and the broader tools for thought community that's been wrestling with how to design systems that extend our cognition, not just record it.

it's about building something that feels less like a database and more like a mind. less like a tool and more like a partner.

lucy's memory isn't perfect. it's not human. but it's designed with these principles in mind because we believe that how an ai remembers is just as important as what it remembers. it's the difference between feeling watched and feeling understood.

if you're curious about how this feels in practice, you can always come talk to lucy yourself.