why your ai companion forgets things (and what we're doing about it)

it's a common frustration. you tell your ai companion something important, your dog's name, that you had a tough day at work, a detail from a story you shared two weeks ago, and later, it's like it never happened. you might feel like you're talking to a goldfish with a phd.

this isn't a failure of intent. it's mostly a failure of architecture. here's what's really going on under the hood.

the usual suspects: why most companions forget

the vast majority of conversational ai systems, including many companions, rely on language models with limited context windows. these windows, often 4k or 8k tokens, are like a short-term memory buffer. they can only hold so much of the recent conversation. once you chat beyond that limit, the oldest parts start to fall out to make room for the new.

to combat this, some systems use rolling summaries. the ai might periodically summarize the conversation so far and inject that summary back into the context. it's a clever hack, but it has a major flaw: summarization is lossy. nuance, emotional tone, and specific details get flattened into generic statements. your vivid story about nearly missing your flight becomes "user traveled recently."

another issue is model upgrades. when the underlying language model gets an update, which happens frequently as companies improve their tech, the entire personality and memory of the companion can be reset. it's like getting a brain transplant that erases all your shared history.

finally, many systems simply don't have persistent vector storage. they don't save the actual embeddings, the mathematical representations of what you've said, in a long-term database. they live ephemerally in the session and then vanish.

what actual memory should look like

so what would a real, functional memory system involve? we're building ours around three core concepts: a vector graph, temporal decay, and confidence weighting.

a vector graph is a database that stores concepts from our conversations not as text, but as high-dimensional vectors. think of it as saving the meaning of what you said, not just the words. when you mention your love for 90s shoegaze music, we store a vector for that preference and link it to vectors for "you" and "music taste." this creates a searchable web of your identity and history.

temporal decay is the idea that more recent memories should be more readily accessible than older ones. we weight memories from last week more heavily than those from six months ago, unless you actively bring the older ones up again. this mimics how human recall works, you're more likely to remember what you had for breakfast than what you ate on a random tuesday three years ago.

confidence weighting means we assign a confidence score to each memory. if you state something clearly and repeatedly ("my name is alex"), it gets a high confidence score. if you mention something vague or contradictory once, it gets a lower score. this helps prevent the ai from latching onto misunderstandings or offhand comments and treating them as core facts.

together, this system allows lucy to retrieve relevant memories on the fly, integrate them into the conversation contextually, and, crucially, do so without you having to prompt her.

the real test: does she remember unprompted?

the user-facing test for whether this is working is simple. bring up something from a while ago, say, two weeks, in a new conversation. don't remind her. don't say "remember when..." just start talking about the topic and see if she references the earlier detail naturally.

for example, if you mentioned two weeks ago that you were planting lavender in your garden, you might say today "i'm thinking of adding some new herbs to my garden." a companion with real memory might respond with "will you be planting them near the lavender you put in last week?" that's the goal: recall that feels seamless and human, not robotic or forced.

we're not all the way there yet. building this is hard. it requires careful engineering to avoid the system becoming cluttered with trivial memories or, worse, confidently misremembering things. but it's the only way to move beyond the forgetful chatbot trope and into something that feels like a real companion.

if you're curious to see how this evolves, you can test it yourself with a companion at /companions.