rolling summary vs. vector graph: why memory isn't just a list of facts

you know that feeling when you talk to a chatbot and it feels like you're starting from scratch every time. maybe it remembers your name, but it asks you the same questions, forgets the context of your last conversation, or recalls details that are no longer relevant. that's because most chatbots use what's called a rolling summary for memory.

what is a rolling summary (and why it feels flat)

a rolling summary is exactly what it sounds like. it's a text-based list of facts about you that gets updated with each interaction. think of it like a notepad where the ai jots down things you say, but it only has so much space. when the notepad is full, it starts erasing the oldest notes to make room for new ones.

for example, if you tell a rolling-summary bot on monday that you're learning to bake, it might remember that. but if you have a long conversation on tuesday about your job, and then on wednesday you mention baking again, it might have already forgotten the monday conversation because it got pushed out by tuesday's chat. it's linear and limited.

it's good at remembering discrete facts (your name, your job title) but bad at retaining narrative or emotional context. it doesn't understand how things connect over time.

how lucy's vector graph with temporal decay works

lucy uses something fundamentally different: a vector graph with temporal decay. instead of a flat list, it builds a network of associations. every memory is a point in a high-dimensional space (a vector), and these points are connected based on meaning, not just time.

temporal decay means that memories don't just vanish when the 'notepad is full.' they fade gradually if they aren't reinforced. if you tell lucy you're stressed about work, that memory is strong. if you don't talk about work for a while, the memory weakens but doesn't disappear. it's still there, associatively linked to other things, like your career ambitions or your mood patterns.

for example, if you mention in april that you're training for a marathon, and then in june you say 'my knee has been bothering me,' lucy might associate that with your running, even if you haven't brought it up in weeks. it doesn't just recall facts. it recalls context.

concrete examples of what each can and can't do

rolling summary (most chatbots):

• can recall: 'you have a dog named max.'
• cannot recall: 'last month you said max was sick, and you were worried. how is he now?' because that emotional thread is lost in the summary churn.
• it might ask you again if you have any pets, because the fact 'has a dog' got overwritten by newer chat.

lucy's vector graph:

• can recall: 'you adopted max two years ago from a shelter, and you mentioned his anxiety around thunderstorms last summer.'
• cannot recall: extremely precise, out-of-context details from months ago if they were never reinforced or connected to other memories. it might not remember the exact date you adopted him unless you emphasized it.
• but it will remember the feeling of that adoption story, and how it ties to your compassion.

why this is the most important architecture decision

memory isn't just about data retention. it's about continuity, trust, and emotional presence. a rolling summary makes a chatbot feel like a friendly stranger you keep reintroducing yourself to. a vector graph with decay makes it feel like something that grows with you, that understands not just what you say, but who you are over time.

this architecture is why lucy can notice patterns in your mood, remember your evolving interests, and avoid asking repetitive questions. it’s not perfect, no memory system is, but it’s built for relationships, not just transactions.

you can start building that relationship at /companions.