The rule. When an AI product remembers users across sessions, ship standing controls with it: a place to see what is stored, a way to edit or delete it, visibility into when it is used, and a global off switch.
Why. Guideline 17 of the human-AI interaction guidelines, in the set's over time phase, states the requirement plainly: allow the user to globally customize what the AI system monitors and how it behaves (Amershi et al., 2019; Microsoft, n.d.). The flagship assistants have converged on the same control anatomy: ChatGPT pairs memory with an editable memory summary, a per-response sources view, temporary chats, and an off switch (OpenAI, n.d.); Claude ships memory fully optional, with an editable summary, a separate memory per project, and incognito chats that never write to memory (Anthropic, 2025). The candid footnote in OpenAI's own FAQ, that the summary "will not include everything that ChatGPT remembers," marks exactly where surprises live: whatever the controls do not surface, the next reply might (OpenAI, n.d.).
Seen in the wild. ChatGPT's sources view puts a book icon under a personalized response; tapping it shows which memories shaped the answer and why each one was used (OpenAI, n.d.).