An assistant correctly recalls a user's tree-nut allergy. Given a request for macarons, it supplies an almond-flour recipe without applying that allergy to the user.
This is a recorded example from InMind, a July 2026 study of agent memory, using xMemory. The evaluations ask two different things: can the system retrieve a fact when the question names it, and can it bring that fact into a decision that needs it?
The second question is the reason to give an assistant memory in the first place. A user should not need to know which past conversation to reference before every request.
For a coding agent, the equivalent might be remembering that a service runs in short-lived processes, then proposing an in-process timer for hourly cleanup. That is an illustrative test case, not a result reported in the paper. It has the same useful structure: the current request does not repeat the architectural constraint, but the constraint changes what a correct solution looks like.
InMind makes that distinction measurable. Its 125 synthetic tasks pair a stored personal fact with a later request whose relevance depends on background knowledge. The authors deliberately remove obvious lexical and semantic retrieval cues. This is a stress test for a particular failure mode, not a sample of everyday assistant traffic.
