AI Memory Doesn't Work the Way You Think

People keep talking about adding memory to LLMs like it's just another layer. But the second you try to "remember" stuff across sessions, things fall apart.

The model doesn't know who you are. It doesn't know who it is. If you want continuity, you need a wrapper that saves info somewhere and injects it back in later. But that's brittle. It's weird. It's not really memory. It's document injection dressed up as memory.

Real memory would mean state. And that's a whole other problem.