How do AI products decide what to remember between conversations?

Nielsen Norman Group's "Artificial Intelligence" glossary (Caleb Sponheim, 2026-08-21) draws a specific line between two terms that are often used interchangeably. "Context" is defined as "everything supplied or made available to a model for a single request: the prompt, the system prompt, selected conversation history, retrieved documents, stored memory, and the results of any tools it has called." "Memory" is defined separately as "information an AI product carries forward in subsequent interactions with a user," split further into "conversation, or short-term, memory" that "retains recent interaction within a session" and "long-term memory" that "consists of information stored across sessions and retrieved when relevant." The glossary states the relationship directly: "Context is what is available for one request; memory is retained information that may later become part of that context."

The mechanism the glossary describes is assembly, not storage. Context is not a single thing the model holds onto; it is built fresh for each request out of several contributors — prompt, system prompt, history, retrieved documents, memory, tool results — and memory is only one of those contributors, promoted into a given request's context when "retrieved when relevant." The glossary attributes this assembly work to the product layer, not the model: "In most current assistant products, the surrounding application or service stores conversational state and supplies or references it when relevant." Naming this as a deliberate design choice is what the glossary calls "context engineering" — "choosing deliberately what goes into the context."

This sharpens How does AI context differ from conventional software context?, which argues that context is a model's reconstructed-every-turn working window rather than a fixed substrate like a conventional UI. That note treats context as a single mutable thing; NN/g's glossary names the specific mechanism inside that mutability — memory, split into short-term and long-term — that supplies some of what gets reconstructed into context at each turn, and makes explicit that the reconstruction is a product-layer decision rather than a property of the model itself.

The glossary names the pieces — prompt, history, retrieved documents, short-term and long-term memory, tool results — but does not say how a product decides which long-term memory items to surface into a given request, or how a conflict between stored memory and the live conversation gets resolved. It states that context engineering is "choosing deliberately what goes into the context" without describing what that choosing looks like in practice. The implication that follows is narrower than it looks: because the context/memory boundary is a product-layer decision rather than a fixed property of the underlying model, the same model can appear to remember very differently across products depending on how each one retrieves and inserts memory into context — which means claims about what an AI remembers are claims about the product, not the model.