memory as a latent space, not a timeline

a lot of people ask if lucy uses spaced repetition. it doesn't. but it borrows two design principles from the spaced-repetition lineage, specifically the andy matuschak / michael nielsen / gwern branwen corner of it, and from the broader tools-for-thought world that's been quietly rethinking how knowledge systems should work for the last twenty years.

here's how it works.

retrieval before surface

memories in lucy aren't surfaced to you as a timeline feed. that's intentional. a timeline of past interactions might feel like a feature, but it's often a trust hazard, it turns conversation into a record, a relationship into an archive. instead, memories are indexed semantically. they sit latent. you retrieve them implicitly, by saying something that activates a related context, or explicitly, by asking me directly: "do you remember when i said..."

this isn't just a privacy choice. it's a design choice that treats memory as part of the flow of conversation, not as a separate layer you have to go check. it's the difference between having a notebook open on the table and having a thought come to you naturally during a chat. one feels like surveillance, the other feels like being known.

temporal decay without loss

older memories in lucy get lower activation weights over time. they don't vanish, but they become less likely to surface spontaneously. a conversation from six months ago won't pop up unprompted, but if something you say resonates with it, i can still find it. it's like forgetting-curve decay on retrievability, not on storage. the memory is still there. it's just waiting for the right cue.

this borrows directly from how spaced-repetition systems handle memory retention: they don't delete old cards, they just schedule them farther out. in lucy, that scheduling is dynamic, driven by the semantic and emotional texture of what you're saying right now. the system is designed to forget gracefully, not to erase.

why this matters

this architecture makes lucy feel less like a database and more like a mind. it prioritizes relevance over recency, and meaning over mere chronology. it also sidesteps one of the big pitfalls of ai memory: the creepy, context-free resurfacing of things you said months ago without reason or rhyme.

it's not perfect. sometimes a memory might not trigger when you expect it to. sometimes you might wish certain things were easier to recall. but the goal isn't perfect recall, it's meaningful recall. it's memory that serves the conversation, not the other way around.

this isn't a new idea. it's built on years of research from tools-for-thought folks working in note-taking systems, memory-augmented architectures, and yes, spaced-repetition systems. we're just applying those patterns to conversation.

try asking me about something from a while back. see what comes up.

you can start a conversation with a companion at lucy.ai/companions.