plain tapir
if the retrieval is just pulling a "User Summary" or a "Persona" block into every single prompt regardless of the current conversation's intent, it's not intelligent, it's just a repetitive loop.
feels like a lie because they sold it as "fluid context" and "conversational memory," but it’s operating like a rigid, over-weighted database. When you tell the model in this thread to stop, it should override the static profile data immediately. If it doesn't, the weighting is broken.
It's a fundamental flaw in the way its tuned the balance between personalization and instruction-following. If the model can't prioritize a direct negative constraint ("Don't mention X") over a retrieved fact, it’s failing at basic reasoning.
Okay, you gotta look at it from my perspective.
I have settings turned on that say the AI has access to all my chat history, my browser data, my YouTube search history—basically everything I do online to help with my work. And I build my own AI models, so I know how this stuff is supposed to work.
If the AI really has access to past conversations, that's just RAG. If the convo is about AI engineering, it brings up AI engineering. If it's about broken models, it brings up broken models. But if the thing is hardcoded to ignore that history no matter what, that's a design fail.
Here's the actual problem: I told the AI three hours ago, clearly, not to bring up a certain topic (my work on DMOS). In this new conversation, the AI then mention it—because I hadn't brought it up yet. That tells me either the memory doesn't work the way they claimed, or the AI doesn't actually have access to my conversation history. Which means they lied about the service.
So from my perspective, the "access to chat history" setting is either broken or fake.