An assistant remembers that you prefer short replies. That sounds useful until you wonder where the preference came from, which conversations it applies to and how to change it.
A useful memory design needs a source. The user should be able to distinguish a preference they explicitly selected from a pattern the system inferred. Those are different kinds of evidence.
It also needs a scope. Short replies may fit routine work updates while being inappropriate for a personal conversation. A memory should not silently become a rule for every relationship.
Finally, it needs correction. People change their minds, and systems misunderstand examples. ‘Use more detail with this person’ should update the relevant behaviour rather than require the user to rebuild their whole profile.
Removal matters too. Ask what deletion means for the current product and whether it affects the information used in later suggestions. Do not assume that a visible remove button answers every data-handling question.
When considering a personal AI tool, inspect these three practical areas: source, scope and change. A larger memory is not automatically a better one. Its value depends on whether the information is relevant, understandable and still appropriate to use for the decision being made now.