Tangent: Annoyed at the GPT-ish illustration style on this article because I see it everywhere, including our site. Not sure if the model styles have homogenized (training on AI images, model collapse) or we all just stopped being detailed enough in our image prompts to get diverse results.
key graf:
Bots are docs: LLMs can’t keep secrets, no matter how well prompted, we need to store memories in a secure, separated way. Fortunately, we already have a good mechanism for this kind of semi-shared secret: documents. Asking an agent to only disclose some of what it knows about you is a bit like telling someone only to read one page out of a long document that has all of your life in it - you’d never do that! We know how to separate and share documents.
I hear there's a new startup named Rekall that are doing memory implants...