https://twitter.com/doesdatmaksense

We’re still early when it comes to AI memory and personalization. After spending past few months building and working with memory systemss, it’s clear how far we still have to go to even get close to human-level memory. Most current systems just extract what looks important from a conversation or context and store them as isolated bits the model can refer back later to personalize responses or behavior. That works for basic personalization, but it still largely depends on chat-like interactions or text context for the memory to form and update in meaningful ways.

That’s very different from how human memory works. We build memory from all kinds of inputs: what we see, hear, read, say, feel. Some memories stick because they’re emotional. Others just repeat enough times that we learn them. What we remember is shaped by everything we’ve experienced, not just what was explicitly said or done.

Key difference is that humans don’t always need structured interactions to learn or remember. We observe and infer, and then form memories passively even through things we never consciously pay attention to.

Where this is heading, I think, is toward memory that doesn’t rely on conversation as the main channel. Tools are starting to pick up signals from other modalities. Vision-based memory is something I feel will be emerging in the coming months. Models will begin to “see” what users see whether it’s during browsing, shopping, working, or learning, and start to build memory from behavioral signals, attention holding patterns, and environmental context. For example, what made someone click a particular link, hesitate on a page, or return to a product multiple times. These can shape much deeper personalization than conversation alone.

Pretty exciting times ahead!