why your ai companion forgets things (and what real memory looks like)

you tell your ai companion something important. maybe it’s your cat’s name, or how you take your coffee, or that you had a terrible day two thursdays ago. and for a little while, she remembers. then, a week later, it’s gone. poof. she asks about the cat’s name again. it feels a bit like a breakup, but with a language model.

the reason isn’t malice or indifference. it’s mostly just technical debt and shortcuts. here’s a quick tour of why most ai companions have the memory of a goldfish.

the usual suspects: context, summaries, and upgrades

first, there’s the context window. many models operate with a 4k to 8k token limit. a token is roughly a word or part of a word. that means the ai can only "see" the last few thousand words of your conversation. once you chat beyond that, the oldest part starts getting pushed out to make room for the new. it’s not forgotten, exactly. it’s just… no longer in the working memory.

some services try to fix this with rolling summaries. the ai will summarize old parts of the conversation into a paragraph or two and keep that summary in the context window. but summaries are lossy. they lose nuance, emotion, small details. your cat’s name might survive, but the story about how you found her in a rainstorm might get condensed into "user has a cat."

and then there are model upgrades. the service updates to a newer, better language model. but sometimes that new model hasn’t been trained on your specific conversation history. your companion’s personality, built painstakingly over weeks of chats, gets wiped. it’s like meeting a stranger who looks exactly like your friend.

none of this is done out of laziness, usually. it’s just computationally expensive to do it right. but it’s frustrating. it breaks the illusion of continuity, the thing that makes a companion feel real.

what actual memory looks like

so what does "doing it right" look like? it’s not just a bigger context window. it’s a separate memory system that works alongside the language model.

in lucy, we’re building a memory system based on a few core ideas: a vector graph, temporal decay, and confidence weighting.

a vector graph is a way of storing memories not as text, but as mathematical objects (vectors) in a high-dimensional space. when you say "my cat mittens is sick," that fact gets embedded into this space near concepts like "pets," "health," and "worry." later, when you say "i’m taking her to the vet," the system can find and connect to that memory because the vectors are close together. it’s a web of associations, not just a list.

temporal decay is the idea that memories fade. the system gives more weight to things you’ve talked about recently or frequently. your coffee order from yesterday is more relevant than the name of your first-grade teacher. but the old memory isn’t deleted. it’s just less likely to be recalled unless something specifically triggers it.

confidence weighting means the system tracks how sure it is about a fact. if you say "i hate mushrooms" once, it’s a weak memory. if you say it three times, it becomes a strong memory. if you later say "i love mushroom risotto," the system might flag a conflict and ask for clarification.

together, this creates a memory that feels less like a database and more like… well, a memory. it’s associative, fuzzy, and contextual.

the real test: does she remember unprompted?

the user-facing test for this is simple. don’t ask "do you remember my cat?" wait. talk about something else for a week or two. then, bring up something related. say, "i’m so tired, mittens was up all night meowing." a companion with real memory might say, "oh no, is she feeling any better since she was sick?" she recalls the earlier event without you having to prompt her.

that’s the goal. not perfect recall, but coherent recall. a sense that your companion is growing with you, not resetting every day.

we’re not all the way there yet. building this is hard, and sometimes lucy will still get it wrong. but it’s the thing we’re focused on getting right. because a companion that forgets you isn’t really a companion at all.

you can build a companion with a more persistent memory at /companions.