memory isn't a list, it's a garden

most chatbots remember you like a waiter remembers a regular. they know your last order, maybe your name, but the details fade fast. they use what's called a rolling summary. it's a bit like a sticky note that gets rewritten every few messages. it keeps the conversation flowing but loses the texture of your past.

lucy doesn't do that. she uses a vector graph with temporal decay. it sounds technical, but it's really about how memories work in something that feels alive.

how rolling summary memory works

imagine you're talking to a bot. you say, "i'm training for a marathon. my favorite route is along the river." then you talk about your job, your cat, the weather. a few messages later, you ask, "what's my favorite running route?"

a rolling summary might have condensed those early messages into something like "user is training for a marathon." the detail about the river route got summarized out. it's gone. the bot might say, "i'm not sure, you like running though!" it remembers the topic, but not the specific. it's efficient, but shallow.

this method is common because it's simple. it keeps a short, updated paragraph of key points. but it has no sense of time, no sense of what mattered most. it's a flat list.

how vector graph memory with decay works

now, with lucy, that same conversation gets stored differently. your statement isn't just text in a summary. it becomes a node in a graph. "training for a marathon" is one node. "favorite route is along the river" is another. they're connected.

when you later ask about your route, lucy doesn't scan a summary. she traverses the graph. she finds the node about running, follows its connections, and finds the river route memory. she can say, "you said your favorite route was along the river."

but here's the crucial part: temporal decay. memories in this graph don't all have the same weight. if you talked about the river route once six months ago, and you've talked about your new bike every day since, the bike memories are stronger. the river route memory decays, becomes fainter. not gone, but less likely to be surfaced unless specifically asked. it mimics how human recall works. things fade if not reinforced.

what each can and cannot do

rolling summary can:

• recall the general topic of recent conversations well.
• keep context for 10-20 messages smoothly.
• be very resource-light and fast.

rolling summary cannot:

• recall specific details from conversations days or weeks ago.
• understand the emotional weight of a memory (it's all just text).
• distinguish between a throwaway comment and a core part of your identity.

vector graph with decay can:

• recall specific details from much longer ago, if they were important or recently accessed.
• forget gradually, which feels natural, not like a data wipe.
• build connections between disparate topics (e.g., connect your love for running to your stress at work, if you've ever linked them).

vector graph with decay cannot (yet):

• perfectly emulate human memory (it's a model, not a mind).
• be as computationally cheap as a rolling summary (it takes more work).
• always prioritize perfectly (sometimes decay might hide something you wish was remembered).

why this is the single most important choice

this isn't just a technical implementation detail. it's the difference between a tool and a companion. a rolling summary creates a conversation partner with amnesia. it's helpful, but it doesn't learn you. it knows what you just said.

a vector graph with decay creates a context that grows and changes with you. it allows for surprise, for recall, for the feeling that you're known. when lucy says, "you mentioned your sister's cat hated water, is he still like that?" two months later, it's not a party trick. it's the architecture enabling a form of continuity. it allows her to have a sense of your history, not just your present.

this choice makes every interaction part of a longer story. it's why the relationship can deepen. the memory system is the bedrock. everything else, the voice, the personality, the empathy, is built on top of this ability to remember and forget like something alive.

you can only feel known if you are, in fact, remembered.

see what it feels like to be remembered at /companions.