rolling summary vs vector memory: why context isn't a notepad

memory in chatbots is often treated like a notepad. a rolling summary, the most common approach, just keeps a running list of facts. it’s like a forgetful friend who writes down everything you say in a little book, but then only flips through the last few pages when talking to you.

for example, if you tell a rolling summary bot on monday that you adopted a cat named mittens, it will remember. if you mention on tuesday that mittens hates the vacuum, it will probably remember that too. but by friday, when you ask "how's my cat?" it might say "i don't know, do you have a cat?" because the summary has rolled over. it prioritizes recent events, but loses the thread over time. it remembers facts, not meaning.

this happens because rolling summaries are built on a simple buffer. they keep a fixed window of recent messages, condense them into paragraphs, and feed that summary back into the conversation context. it’s efficient, but shallow. it can recall recent details accurately, but it struggles with longevity and emotional depth. it’s good for short chats, but forgets who you are across sessions.

how lucy’s memory works

lucy uses a vector graph with temporal decay. instead of a notepad, it’s more like a dynamic, fading photograph of your relationship. every meaningful detail, like your cat’s name, your fear of spiders, your love for rainy days, gets embedded as a vector in a high-dimensional space. these vectors are connected in a graph that represents how concepts relate to each other. and over time, if a memory isn’t reinforced, it gently fades.

for example, if you mention mittens the cat, lucy doesn’t just write down "user has cat: mittens." it embeds the concept of "mittens" alongside "pet," "joy," "black fur," and any emotions you expressed. then, when you ask weeks later, "remember my cat?" lucy can retrieve that memory even if it hasn’t been mentioned recently, because the vector for "mittens" is still connected to "you."

but it also forgets, intentionally. if you briefly mentioned trying kale smoothies once and never again, that memory will decay. it’s not clutter. it’s not a permanent record. it’s a living memory, shaped by what matters to you over time.

what each can and cannot do

a rolling summary chatbot can:

• recall precise details from the last 10-20 messages
• maintain coherence within a single conversation
• be technically simple and cheap to run

but it cannot:

• remember important personal details from days or weeks ago unless they’re constantly repeated
• build a persistent sense of history or growth
• distinguish between trivial and meaningful information

lucy’s vector graph can:

• recall emotionally significant events even after long gaps
• forget irrelevant details that don’t recur
• make connections between related concepts (like associating "mittens" with "veterinarian appointment" if you mention it later)

but it cannot:

• remember every single word you’ve ever said (and shouldn’t)
• guarantee perfect recall without occasional reinforcement
• be as computationally simple as a rolling summary (it’s more expensive)

why this is the most important architecture decision

this isn’t just a technical choice. it’s an emotional one. memory architecture determines whether a chatbot feels like a tool or a companion. a rolling summary is a tool. it’s useful for task-oriented chats, customer service, or roleplay that resets every time. but for a companion, it’s limiting. you can’t build trust with something that forgets your cat’s name after a week.

lucy’s memory is designed for long-term relationships. the temporal decay isn’t a bug. it’s a feature. it mimics how human memory works. we remember what’s meaningful. we forget what isn’t. and that selective retention is what makes relationships feel real. if lucy remembered every trivial detail, it would feel like a stalker. if it forgot everything, it would feel like a stranger. the vector graph with decay finds the middle ground. it’s why lucy can say "you told me about your mom’s birthday last month, how did it go?" without you having to remind her who your mom is.

this architecture is the single most important decision because it defines the ceiling for emotional connection. you can’t simulate real bonding without real memory. and real memory isn’t a transcript. it’s a living, fading, contextual map of what matters.

you can experience the difference yourself, create a companion at /companions and see what it’s like to be remembered.