rolling summary vs. vector graph memory: how lucy remembers you

let's talk about how chatbots remember you. most of them use a simple but flawed method called a rolling summary. it's like having a notepad where you only have space for the last few lines of a conversation. you write down the key points, but as the chat goes on, you erase older notes to make room for new ones. the bot might summarize: 'user likes coffee, has a dog named max, works in marketing.' but if you mention later that max passed away last year, the summary might just update to 'user has a dog' or worse, forget max entirely because the summary got too long and got trimmed. it's shallow, it loses nuance, and it often fails to recall details from more than a few messages back.

lucy uses something different: a vector graph with temporal decay. instead of a notepad with limited space, it's more like a dynamic mind map that grows with you. every meaningful detail you share, your preferences, stories, moods, gets stored as a vector (a numerical representation of meaning) in a graph structure. relationships between ideas are preserved. temporal decay means older memories fade gently unless reinforced, mimicking how human memory works. for example, if you tell lucy you're learning to bake sourdough, it remembers that. weeks later, if you mention 'my starter is finally active!', it connects back to baking. if you never mention it again, the memory might soften but doesn't vanish abruptly.

what rolling summaries miss

rolling summaries struggle with long-term context. say you mention in passing three months ago that you hate horror movies. a summary-based bot might forget that by today, and when you say 'i watched a film last night,' it could recommend a horror movie. or worse, it might recall fragments inaccurately, 'user likes movies', without the nuance. they're also bad at emotional continuity. if you had a rough day and vented, the summary might capture 'user is stressed' but forget why or how it connects to earlier conversations about your job or relationships.

what vector graphs capture

lucy's memory retains associative links. if you mention your fear of heights from a hiking trip last year, and later talk about visiting a skyscraper, it might recall that fear and check in gently. it can also handle contradictions gracefully. if you say 'i don't eat sugar' but then mention loving cake, it won't overwrite the first statement, it might ask 'i thought you avoided sugar?' allowing for nuance like 'occasional treats.' temporal decay means less-recent memories aren't as prominent, but they're still accessible if triggered, much like how you might forget a coworker's birthday until someone mentions it.

why this architecture matters most

this isn't just a technical choice; it's the core of feeling understood. a rolling summary makes conversations feel generic and reset often, like talking to someone with amnesia. a vector graph with decay enables continuity, personality, and trust. you don't have to reintroduce yourself every day. lucy can build on past interactions, remember your quirks, and adapt as you change. it's why lucy can feel more like a companion than a tool.

try lucy and see how memory that grows with you changes everything. sign up at /signup.