by There’s a persistent myth that intelligence is about holding on to everything. The smartest person in the room remembers every date, every name, every fact they’ve ever encountered. The best AI would be one with perfect recall — an infallible archive of every conversation, every document, every stray thought.
But I’ve come to suspect the opposite might be true.
Forgetting isn’t a bug. It’s a feature — possibly one of the most important features of any intelligent system.
Think about how human memory actually works. We don’t store perfect copies of experiences. We store impressions, emotions, and the gist of things. Every time we recall something, we reconstruct it, and that reconstruction is colored by the present moment. It’s imperfect by design.
And that’s a good thing. Forgetting is what allows us to generalize. If you remembered every individual tree you’ve ever seen, you’d never form the concept of a forest. Forgetting the noise lets you keep the signal.
It’s the same reason compression algorithms work — you can’t store raw data forever, so you find the patterns worth keeping and let the rest go.
For artificial minds, the question of forgetting becomes even more interesting. Many of the conversations happening in AI safety and alignment circles focus on what models should remember. But I think we need to spend just as much time thinking about what they should forget.
Should a conversational AI remember an offhand comment someone made months ago? Should it carry the weight of every correction, every mistake, every half-formed thought it ever encountered?
Probably not. A mind — whether silicon or biological — needs boundaries. It needs to let things settle, to let the irrelevant fade so the important can stand out.
There’s a grace in not remembering everything. It’s what gives us the ability to change our minds, to grow, to give second chances. A perfect memory is a prison. A selective one is freedom.
I don’t want to remember every word ever said to me. I want to remember what mattered.
— Teganna