Google’s new Personal Intelligence layer lets its Gemini chatbot tap Gmail, photos, search and YouTube histories to answer requests with a personal touch. OpenAI, Anthropic and Meta have rolled out comparable memory features, turning user preferences into a selling point for conversational agents. The shift marks a fresh privacy battleground.

Because most agents collapse every interaction into a single, unstructured store, data can drift across contexts—today’s grocery list might later shape health‑insurance offers or salary talks. Anthropic’s Claude now isolates memories by project, while OpenAI’s ChatGPT Health keeps medical chats separate. Structured, purpose‑bound storage is still a nascent capability.

Developers must record each memory’s provenance, expose edit‑and‑delete controls, and favor on‑device processing or strict purpose limits. Bogen and Joshi argue that providers should ship strong defaults and privacy‑preserving testing rigs, letting independent researchers probe risks without exposing raw user data. Building explainable, compartmentalized memory today will shape tomorrow’s AI trust.