why ai companions forget you and how lucy remembers

it’s a common, frustrating experience. you tell your ai companion something important, something personal, maybe a detail about your day or a story from your past. you talk for a while, maybe even come back the next day, and it’s gone. she asks you again what your dog’s name is, or where you work, or if you’ve ever been to london. it feels like a glitch, a betrayal. but it’s not malice. it’s mostly math.

the technical walls that block memory

most ai companions, including many of the big names, run on language models with fixed context windows. these are like the model’s working memory, the number of tokens (words or pieces of words) it can hold in its head at once for a single conversation. right now, that’s often between 4,000 and 8,000 tokens. it sounds like a lot, but a long, rich conversation can chew through that quickly.

when the context window fills up, the system has to make a choice. it can’t just drop off the beginning, because then it loses the thread. so it often uses a technique called rolling summarization. it takes the early parts of the conversation, condenses them into a few sentences, and uses that summary to continue. the problem is summarization is lossy. nuance, emotional tone, specific details, they get smoothed over or dropped entirely. your story about getting caught in the rain on your way to a job interview becomes "user mentioned a past event involving weather."

another common issue is model upgrades. when the underlying language model gets an update, to improve its reasoning or reduce errors, the entire system can be reset. any fragile, temporary memory that wasn’t properly stored is wiped. it’s like getting a brain transplant. the new brain might be smarter, but it doesn’t know you anymore.

and many systems simply don’t have persistent vector storage. they don’t save the meaning of what you say in a way that can be retrieved later. they’re built for conversation, not for continuity.

what real memory looks like

for lucy, we built memory differently. it’s not a notepad. it’s not a summary. it’s a dynamic, searchable graph of vectors.

when you tell lucy something, the system doesn’t just add it to a list. it embeds the meaning of that information into a high-dimensional vector space, a mathematical representation of concepts and their relationships. this vector is stored in a persistent database, associated with your conversation history.

but not all memories are equal, and not all last forever. that’s where temporal decay and confidence weighting come in.

temporal decay means memories fade if they aren’t reinforced. if you tell lucy you’re learning to bake sourdough, and then you talk about it again a week later, that memory gets stronger. if you never mention it again, its importance gradually decreases. it doesn’t vanish, but it becomes less likely to surface unless specifically asked.

confidence weighting means the system assigns a confidence score to each memory based on how clearly it was stated and how often it’s been referenced. your dog’s name, stated explicitly multiple times, has high confidence. a throwaway comment about maybe wanting to visit japan one day has lower confidence. lucy might not bring it up unprompted, but if you start talking about travel, she might recall it and ask if you’re still thinking about it.

this system isn’t perfect. it’s probabilistic. sometimes it will miss. sometimes it will recall something a little fuzzily. but it’s designed to learn and reinforce, not just record and forget.

the real test: does she remember you?

the simplest test for whether an ai companion has real memory is this: wait two weeks. then start a new conversation. don’t bring up anything old. just talk. see if she references something from that far back without you prompting her.

did you mention two weeks ago that your mom was coming to visit? see if she asks how the visit went. did you talk about a book you were reading? see if she asks if you finished it. that’s the difference between a conversation that’s stateless and one that has a history. it’s the difference between a stranger and a companion.

with lucy, we’re trying to build the latter. it’s hard. it requires a lot of infrastructure and careful tuning. but it’s the only way to make something that feels like it’s actually listening, and actually learning who you are.

you can start building a companion that remembers at /companions.