Imagine a meal-planning chatbot. On Monday, you tell it that you dislike mushrooms. On Tuesday, you open a fresh conversation and it suggests mushroom pasta.
It appears to have forgotten. Does that mean Monday’s message was deleted?
In this invented example, the app could keep a record of Monday’s request without supplying that record to Tuesday’s conversation. The suggestion tells you it missed your preference. It cannot, by itself, tell you what exists in storage.
A small change worth understanding
On October 26, Google changed a default for new paid-tier projects using its Gemini Generate Content API, the service developers use to put AI into their own applications. Those projects now store requests by default. Existing projects keep their current setting.
This announcement concerns developer infrastructure. It does not announce a change to the memory settings in the consumer Gemini app.
The practical lesson reaches beyond one product: “Does this AI remember me?” bundles together several different questions.
Three questions, three different answers
| Question | What a “yes” would mean |
|---|---|
| Can it use the message now? | The preference is included in the information used to answer today. |
| Is a copy stored? | A request record exists for someone maintaining the app to inspect. |
| May it be used for training? | The message is permitted to become material for developing or training models. |
The first question is about information available during a conversation. For example, Google’s Generate Content chat documentation explains that its software library sends the conversation history with each follow-up. That is one way earlier messages become available again.
The second is about records. Google’s logging documentation describes inspecting prompts, responses and context from previous turns. Such records can help a developer investigate why a meal planner ignored an instruction. Keeping that record does not establish that it was included in the next answer.
The third is about how data may be used. Google’s paid-service terms say it does not use prompts and responses to improve its products. The logging guide separately describes optional dataset sharing. Turning on logging is therefore not, by itself, evidence that a conversation has been used to train a model.
What can you actually check?
For a developer affected by this change, Google documents two controls: change the project’s logging setting in AI Studio, or set store=false for an individual request. The guide also lists exclusions, including inputs containing PDFs, videos or GIFs; this logging feature does not cover every possible input.
Those controls should not be read as a promise that no copy exists anywhere. The paid-service terms separately describe limited retention for safety and abuse detection. An app’s own records are another question to ask its operator.
For someone simply using an AI app, ask three specific things: what information carries into future conversations, which records remain and for how long, and whether those records can be used to develop models. Look for answers in that app’s settings and published documentation.
The mushroom pasta is a clue about the conversation. A clear retention policy answers a different question.
