The same AI can seem to remember you perfectly in one app and forget everything in another — why?

This explores whether the same model can seem to remember differently depending on the product it's built into. The corpus says yes, and the reason is simple: in most cases memory isn't part of the model. It's a design choice made in the product. Nielsen Norman Group's glossary separates *context* (what the model sees on a given request) from *memory* (stored information that the product may or may not pull into that context). Because each product decides which memories to store, which to retrieve, and when to include them, a statement like 'the AI remembers X' describes the product, not the model How do AI products decide what to remember between conversations?. Two apps running the same model, with different retrieval rules, will look like two different minds.

Users mostly can't see this. Interviews with ChatGPT users found four incompatible mental models of how memory works: some think it lasts only for one session, some think it's part of the training data, some think the model is actively processing what they tell it, and some have no idea What mental models do users hold about LLM memory?. Those beliefs shape what people choose to share and how much they worry about privacy. So when the same model behaves differently from one product to the next, users have little to go on in explaining why. A related note argues that agents behaving inconsistently, without saying so, stop users from ever building an accurate picture of the system. Variation needs to be reduced or made visible How does inconsistent agent behavior affect user understanding?.

The less obvious point is that even when two products retrieve the same memory, they may not *use* it the same way. Paired 'know' and 'act' tests across 16 systems found that agents often recall a user preference correctly and then fail to act on it. Most of these failures came from misinterpreting the preference, not from failing to retrieve it Why do LLM agents remember preferences but not act on them?. Other work goes further: memories that are accurate and relevant can still make reasoning worse. All five memory frameworks tested scored below a no-memory baseline Can relevant memories actually harm LLM reasoning?. So 'remembering differently' happens at three layers: what the product stores, what it retrieves into the prompt, and how the model handles that material once it's there.

The practical takeaway: if you're comparing AI products, or deciding what to tell one, ask about the product's memory design, not the model's name. Which model is underneath tells you surprisingly little about what will be remembered, or whether remembering will actually help.