# Google Splits Gemini’s Chat Memory From Its Training Pipeline

> Source: <https://rews.cc/a/google-splits-gemini-s-chat-memory-from-its-training-pipelin-f20ff9>
> Published: 2026-11-17 16:50:10+00:00

Until November 16, 2026, Gemini had one privacy switch, Keep Activity, and it did two unrelated jobs at once: it decided whether you could scroll back through old chats, and it decided whether Google could feed those same chats into the pipeline that trains its models. Turn it off and you lost your transcripts. Leave it on and every prompt, attachment and tool call you sent was eligible for training, with no way to keep the first without accepting the second. Google is now splitting that switch into two, as [Ars Technica reported](https://arstechnica.com/ai/2026/11/google-will-let-you-opt-out-of-ai-training-in-gemini-without-losing-your-history/), rolling the change out globally over the coming weeks.

## Two switches, two jobs

The new settings live in the same place the old one did: Gemini app, Settings & help, Activity, or via a Google Account at Data and Privacy, My Activity, Other activity, Gemini Apps Activity. One toggle, History, controls retention length and nothing else. Leave it on and Google auto-deletes your activity on a timer you can set to 3, 18 (the default), or 36 months, or never, according to Google’s Gemini Apps Activity help page. Turn History off and, rather than disappearing, your chats now persist for roughly two months, used, in Google’s words, to “respond to you, process your feedback, and protect Google, its users and the public,” per the same help page and [9to5Google’s reporting](https://9to5google.com/2026/11/16/gemini-history-ai-training/) on the rollout. That is a longer tail than the arrangement it replaces: 9to5Google reports the pre-change retention for History-off chats topped out at 24 hours for responding to you and 72 hours for safety purposes, meaning the floor on how long a “deleted” conversation sticks around just moved from one or three days to about sixty, a 20- to 60-fold increase.

The second toggle, Training, is the one that is actually new. It controls whether your text, attachments, tool calls, location data and information from connected apps feed the models, independent of whether you can see those chats again later. It is opt-out: Google trains on your Gemini activity unless you find the switch and turn it off. 9to5Google reports that if a user’s prior Keep Activity setting was off, both new toggles come up off by default; what happens for users who had it on is less consistently reported. Android Authority describes both History and Training as off by default across the board, while Cryptobriefing describes History defaulting on and Training off, a discrepancy that reflects a staggered rollout rather than one documented global default.

**Gemini's two privacy toggles, after the split** How a chat's fate now depends on two independent switches

| From | To | How | 
|---|---|---|
| Your Gemini chat | History toggle |  | 
| Your Gemini chat | Training toggle |  | 
| History toggle | History on *(18-month default auto-delete)* | on | 
| History toggle | History off *(~2-month retention)* | off | 
| Training toggle | Training on *(chat eligible for model training)* | on | 
| Training toggle | Training off *(excluded from training data sets)* | off | 
| Training on *(chat eligible for model training)* | Human review subset *(kept up to 3 years, survives deletion)* | may be selected | 

The chart shows the asymmetry: History governs how long you can see a chat, Training governs whether Google’s models learn from it, and the two now run on separate rails. But one branch escapes both switches. Google’s support documentation states that human reviewers, including service providers, read, annotate and process Gemini Apps conversations to improve quality and train its generative models, and that conversations selected for this review are disconnected from a user’s Google Account and kept for up to three years, whether or not the user later deletes their activity. That three-year figure matches what Ars Technica’s reporting flagged: Google says it may retain data for up to three years to improve its models even after it has been fed in. Opting out of Training stops future chats from entering that review pipeline; it does not reach back into chats already reviewed.

## What the new settings don’t change

The scope is forward-only. Both new toggles apply to interactions from the point you set them, not retroactively. If you have been using Gemini with Keep Activity on since its 2023 launch, that history is already inside Google’s training corpus, and the company has previously said it tries to filter personal information out of that data without guaranteeing it. Submitting feedback on a response also carves a hole in the History-off setting: Google collects the preceding 24 hours of chat context to process that feedback regardless of your retention choice. Audio and Gemini Live video or screenshare data get their own, separate control, outside the History/Training pair. Temporary chats, a distinct mode, are never saved to History and never used for training.

## Where Gemini now sits against ChatGPT and Claude

Gemini is not introducing a concept rivals lack; it is catching up to a decoupling both already made, on different retention terms. OpenAI separated chat history from training years ago: switching off “Improve the model for everyone” in ChatGPT’s Data Controls, per [OpenAI’s own help documentation](https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance), stops future conversations from training the model without touching how long they sit in your history. Anthropic runs the opposite default. Its [announced policy change](https://www.anthropic.com/news/updates-to-our-consumer-terms) made training opt-out rather than opt-in for Claude Free, Pro and Max users as of September 28, 2025: leave the “Help improve Claude” toggle on and conversations are retained for up to five years; turn it off and retention drops to 30 days. None of the three retroactively pulls data out of a model already trained on it.

The table is a reminder that “you can opt out of training” is not one policy but three, with different floors. Gemini’s shortest-retention path, History and Training both off, still beats Claude’s 30-day floor by about a month, but Google’s three-year human-review exception has no stated Anthropic or OpenAI equivalent in the figures each company has published, which makes the comparison partial rather than clean: each firm discloses a different slice of its pipeline, and none publishes what share of conversations actually gets pulled into human review.

Google’s move follows a pattern of coupling unrelated controls behind one switch and then, under pressure, separating them, the same shape as its decision to [tie school Search access to a single AI on/off toggle](https://rews.cc/a/google-s-off-switch-for-school-ai-also-turns-off-search-8c7d90) that this site covered in early November. The open number here is not in any press release: Google has not said what fraction of Gemini’s hundreds of millions of monthly users will actually find the Training toggle, let alone flip it, and the company that trained a frontier model on activity data by default for three years has no particular incentive to make the opt-out loud.
