You flipped the switch you could see. The other one is still on.

You start a fresh chat about something ordinary (a recipe, a bug, a birthday gift), and halfway through, the assistant mentions that you never take a meeting before noon, or repeats a phrase you used exactly once, or leans on a preference you set months ago and never restated. You never brought any of it up in this conversation. Two questions land at the same instant: how does it know that, and where did the thing I told it actually go. Most people only think to ask the first one.

They are two different questions with two different answers, and the line between them just moved. On March 17, 2026, Google made Gemini’s Personal Intelligence free for people in the US, letting the assistant draw on your Gmail, Photos, and YouTube to personalize its replies, a capability that used to sit behind a paid subscription.¹ More assistants now remember more about more people, and almost nobody can say which of the two systems they have actually switched on.

Two separate systems are in play, with two separate switches. Memory decides what your assistant knows about you the next time you open it. Training decides whether your conversations leave your account to help build the next version of the model. Turning one off does nothing to the other, the defaults differ by vendor, and both changed in the last year. Find both switches for the assistant you actually use. If you settled this last year, you may be feeding a model you meant to opt out of while fretting over the toggle that was never the leak.

If you have ever flipped a memory setting off and assumed that also kept your chats out of the next model’s training, this is for you. If you only ever use a locked-down work or enterprise account, skip to the last section: those tiers are usually walled off from training by default.²

Remembering you and learning from you are two different systems

Here is the whole confusion in one image.

Memory is a dossier the assistant keeps about you and reads before it answers. Training is a donation bin: whatever you drop in gets blended into the next model everyone downloads.

The metaphor only carries you so far, so here is the machinery behind it. Memory is a store of facts and preferences tied to your login. The assistant reads it before replying so it can sound like it knows you, and it stays with your account. Training is different in kind: your conversations are included in the dataset used to improve the model itself, and that improved model then ships to every user. One is about what the assistant knows about you. The other is about what the model learns from you.

The danger is that they feel like one setting. People flip memory off and believe they have gone private, when they may not have touched training at all. People opt out of training and assume the assistant will now forget them, when it may still carry a full memory profile. Two switches, wired to nothing in common.

Memory is personal and stays with your account. Training is collective and leaves it. Switching off one has no effect on the other.

Memory is what it knows about you next time

Nora asks Claude to help word a resignation letter in one chat. A week later, in a new chat about her weekend plans, it refers to her “upcoming transition.” Nothing leaked to anyone. The assistant simply kept a note and read it back, which is memory doing exactly its job, and its job is the part that feels like being watched.

Where that switch lives, and whether it starts on, is different in every product. Gemini’s Personal Intelligence is opt-in and stays off until you connect an app.¹ Claude’s memory sits under Settings then Capabilities as two toggles, one for searching past chats, one for remembering context going forward.³ ChatGPT keeps memory under Settings then Personalization, in a spot separate from its training control.⁴ Same feature, three menus, three starting positions.

Google draws the split inside a single product, which is the clearest tell that these are two systems. Google says Personal Intelligence does not train on your Gmail inbox or Photos library, and in the same breath that your Gemini prompts and the model’s responses can still be used for training.¹ Personalizing to you and learning from you are wired apart even there.

That covers the switch that shapes what it knows about you. The other switch decides whether your words leave the building.

Training is whether your words leave your account

On consumer ChatGPT, Free, Plus, and Pro, your conversations are used to improve OpenAI’s models by default, unless you turn it off under Settings then Data Controls in the toggle named “Improve the model for everyone.”² Your saved memories ride along on that same setting, which is exactly how two separate systems end up tangled: the training control reaches into the memory store.⁴

Anthropic went the other way and made it a decision. In 2025 it began asking every Claude Free, Pro, and Max user to choose whether their chats train Claude, with a deadline of October 8, 2025. Allow it and Anthropic can keep your conversations for up to five years; decline and it holds to the existing 30 days and does not train on them.⁵ And Google, as above, can use your Gemini prompts for training even with personalization switched off.¹

Three vendors, three different answers to the same question. That gap is where people get burned: “I probably opted out somewhere” does not travel from the assistant you used last week to the one you opened today.

Training decides whether your conversation becomes raw material for the next model. On some consumer tiers it is on unless you say no, on others you were asked to pick, and the answer is not the same across vendors.

The default you assumed last year is probably expired

Both changes landed inside the last twelve months, Anthropic’s choice in the fall of 2025 and Gemini’s free rollout in March 2026, so a preference you set, or clicked past, a year ago may no longer describe your account.

This is where it stops being only a privacy question and becomes a reliability one. When a personalized answer is wrong, memory makes it harder to tell why. Was it a stale fact the assistant stored about you, or a fresh invention? You cannot attribute the failure if you do not know what is sitting in the memory to begin with. And an assistant that sounds equally sure whether it is reading a real note or making one up is the same trap covered in why AI sounds confident when it is wrong.

Treat “I set this once” as out of date. Re-check both switches whenever a vendor ships something like Gemini going free, because that is usually when a default quietly moves.

Find both switches

• Open your main assistant’s settings and locate two different things: a Memory or Personalization control, and a data or training control. They are almost never in the same menu, and finding one is not finding the other.
• ChatGPT: memory under Settings then Personalization; training under Settings then Data Controls, in “Improve the model for everyone.”⁴²
• Claude: memory under Settings then Capabilities, as two toggles; your training answer is the consumer data choice you were shown in 2025.³⁵
• Gemini: Personal Intelligence is opt-in per connected app, and your prompts can still be used for training even with personalization off, so check both.¹
• For anything you would not want in a training set, use the temporary or incognito chat mode. On ChatGPT it skips memory and training in one move; most major assistants now offer an equivalent.⁴
• If you are still choosing a default assistant, do you need Fable 5 routes that call.

Assistant (as of July 2026)	Memory control (where / how)	Training default	Retention
ChatGPT	Settings then Personalization, kept separate from the training control4	On by default on Free, Plus, and Pro; opt out under Settings then Data Controls, in “Improve the model for everyone”2	Not specified in these sources
Claude	Settings then Capabilities, two toggles (search past chats; remember context going forward)3	Free, Pro, and Max users were asked to choose, deadline October 8, 20255	Up to 5 years if you allow training; 30 days and no training if you decline5
Gemini	Personal Intelligence, opt-in per connected app, off until you connect one1	Prompts can still be used for training even with personalization off1	Not specified in these sources

The reliability question under all of this is plain once the two systems come apart: whose account your words are in, and which model is quietly learning from them. That is a consumer-scale version of the same attribution problem our research desk works on in what makes an AI agent reliable. Two settings menus stand between you and something most people using these assistants never have: a real answer to who is learning from you.

Footnotes

1. MacRumors, “Google’s Personal Intelligence Now Rolling Out to Free Gemini Users in the U.S.” The March 17, 2026 free US rollout, the opt-in / off-by-default status, the Gmail / Photos / YouTube connections, and Google’s note that Personal Intelligence does not train on your inbox or Photos library while your prompts and responses can still be used for training: https://www.macrumors.com/2026/03/17/gemini-personal-intelligence/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

2. OpenAI Help Center, “Data Controls FAQ.” Consumer conversations (Free, Plus, Pro) are used to improve OpenAI’s models by default unless you opt out under Settings then Data Controls in “Improve the model for everyone,” while Business, Enterprise, and Edu content is not used for training by default: https://help.openai.com/en/articles/7730893-data-controls-faq ↩ ↩² ↩³ ↩⁴

3. Anthropic, Claude Help Center, “Use Claude’s chat search and memory to build on previous context.” Memory settings under Settings then Capabilities, the two toggles for searching past chats and remembering context going forward, and pause / reset controls: https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context ↩ ↩² ↩³

4. OpenAI Help Center, “Memory FAQ.” Memory is a separate control from model training, saved memories are governed by the same training setting, memory lives under Settings then Personalization, and Temporary Chat skips both memory and training: https://help.openai.com/en/articles/8590148-memory-faq ↩ ↩² ↩³ ↩⁴ ↩⁵

5. Anthropic, “Updates to Consumer Terms and Privacy Policy.” The 2025 choice for Claude Free, Pro, and Max users on whether chats are used to improve Claude, the October 8, 2025 deadline, and the retention terms (up to five years if allowed, 30 days and no training if not): https://www.anthropic.com/news/updates-to-our-consumer-terms ↩ ↩² ↩³ ↩⁴