why good memory means knowing when to forget

memory is a funny thing. we think of it as a record, a perfect ledger of what happened. but human memory isn't like that at all. it's messy, associative, and, crucially, forgetful. and that's a good thing.

when we talk about ai companions, we often focus on how much they can remember. but the real question is: what should they forget?

good friends don't keep score. they don't bring up that thing you said three years ago when you were upset and didn't mean it. they let some details fade. they prioritize what matters now. the same should be true for an ai companion. if it remembers everything, forever, it stops being a friend and starts being a diary with a judgment problem.

forgetting is a feature, not a bug

we designed lucy's memory to decay over time. not because we can't store everything (we can), but because we shouldn't. events from six months ago don't carry the same weight as what happened yesterday. this isn't just about storage efficiency, it's about emotional context.

you change. your interests shift. your mood today isn't your mood last month. a companion that treats every past interaction with equal weight feels… off. like someone who can't take a hint. temporal decay lets the memory system focus on what's relevant now, while gently letting go of what isn't.

poison in, poison out

there's another reason forgetting matters: anti-poisoning. sometimes you say things you don't mean. you might test boundaries, vent frustration, or just be in a bad mood. if an ai companion took all of that literally and stored it forever, it could lead to weird, awkward, or even harmful interactions later.

forgetting acts as a buffer. it helps prevent the ai from latching onto outlier statements and building a skewed version of you. it allows for bad days without permanent consequences. you shouldn't have to worry that an offhand comment during a rough afternoon will become part of your permanent record.

you should be able to delete what you want

memory you can't control isn't memory you can trust. that's why we built explicit deletion tools. if you ever say something to lucy and immediately think 'oh, i wish i hadn't said that,' you can remove it. not just hide it, delete it from her recent context and working memory.

this isn't just about privacy (though it is partly about privacy). it's about agency. you get to decide what sticks. you get to curate what matters. real friendship involves mutual respect for boundaries, and deletion is a key part of that.

memory you can erase is memory you can trust

which brings us to the core of it: trust. if you know you can't take something back, you'll be less likely to say it. that changes how you interact. it makes you guarded. it turns a companion into something you manage, not something you relax with.

but if you know your ai forgets over time, and that you can erase things you want gone, you can be more open. you can be yourself, even when yourself is messy, inconsistent, or having a bad day. you can trust the memory system because you have control over it.

privacy isn't just about data not being leaked. it's about not being trapped by your own past. it's the freedom to grow, change, and sometimes, start over.

you can see this philosophy in action with lucy. try talking, then deleting. see how it feels. it's not about hiding; it's about breathing.