Google Opens Gemini’s Memory Vault: Users Can Now Audit and Edit What the AI Remembers

Google has begun rolling out a feature that finally cracks open one of the most opaque corners of consumer AI: the personal profile that Gemini builds from your conversations. From the app’s settings, users can now open a memory summary, read what the assistant has concluded about them, and rewrite or delete individual entries. The move marks a meaningful step toward auditability for a technology that has spent years asking users to trust a black box.

  • Full visibility: A bulleted memory summary, sorted into categories such as personal data, career history, and interests, is now accessible from the Personal Intelligence menu.
  • Line-by-line editing: A pencil icon opens an edit mode where users can rewrite or remove any stored line, with a timestamped history of changes retained.
  • Conditional availability: The feature is live in the United States for users on a paid Google AI plan with memory enabled, leaving most global users outside the rollout.

Inside the Memory Summary

The interface is straightforward. On the web version, users open Settings via the gear icon in the lower-left corner, navigate to Personal Intelligence, and select the memory summary. Gemini then presents a categorized list of what it has retained. The edit button in the upper-right corner unlocks a mode where any line can be rewritten or deleted outright, or changes can be requested through a text box. Saving preserves a timestamped edit history, effectively handing users a red pen for their own file.

The summary lives inside the same Personal Intelligence menu that recently gained a connection to Google Wallet for reading expenses, suggesting Google is consolidating everything the assistant knows about a user into a single control surface. That architectural choice matters: it turns scattered memory settings into a destination users can actually find.

The Two-Switch Architecture Behind Memory

Whether Gemini remembers anything at all depends on two chained settings. Keep Activity stores chats in the Google account, and Memory only functions if the first is active. Keep Activity is enabled by default for users over 18 with a personal account; disabling it deletes chats after 72 hours and stops syncing across devices. Memory requires Keep Activity to be on in order to store preferences and personal details.

A toggle in the Personal Intelligence menu can halt memory, but with an important caveat: switching it off stops new memories from being saved without erasing existing ones. Removing those requires visiting the memory summary and deleting entries individually, or wiping conversations from Settings. That asymmetry is the crux of the privacy critique. The design still leans toward accumulation, and exercising control demands deliberate effort.

The transparency is real, but it remains a transparency you have to go looking for, clicking through gears and pencils, while remembering happens on its own.

Google’s broader retention picture adds context. The company recently updated its Gemini Apps Privacy Hub to state that temporary chats and chats made with history off are retained for around two months, up from the previously documented 72 hours. Google said the change was intended to improve safety and security. Separately, the company has been splitting Gemini’s Keep Activity setting into distinct controls for chat history and AI training, with both replacement settings disabled by default for users who had previously turned Keep Activity off. Documentation also describes human review and retention of reviewed conversations for up to three years, and data exchanges between Gemini and connected services can include chat content, location, emails, files, calendar events, photos, and videos. Disconnecting an app or deleting its source data does not automatically remove related Gemini Apps Activity.

Why Auditability Matters More Than the Feature Itself

Personalization has long operated as an opaque promise: the assistant learns about you, but you cannot see its conclusions or correct its mistakes. Making memory inspectable and editable is a maturity milestone rather than a decorative extra, and it sets a benchmark competitors will face pressure to match. ChatGPT, for instance, moved in June 2026 to a continuously rewritten memory summary generated by a background process OpenAI calls dreaming, replacing its older saved-memories list, and its controls still vary by plan, region, platform, and workspace.

The limitations here are real. Availability is restricted to paying users in the United States, excluding the vast majority of Gemini’s audience. The default posture still favors retention, and turning memory off does not purge what has already accumulated. The control exists, but the burden of using it falls on the user, one settings menu at a time. For enterprises and privacy-conscious consumers alike, the lesson is that visibility is not the same as deletion, and an auditable memory is only as useful as the diligence applied to auditing it.

Sources & References