How to See What ChatGPT and Gemini Know About You: Prompts and Settings

How to See What ChatGPT and Gemini Know About You: Prompts and Settings

If you want to know what ChatGPT remembers about you, or what Gemini knows about you from past chats, you need to check two things: the settings panel and the model's own answers. Neither gives you the full picture on its own. This guide walks through exactly how to see what ChatGPT and Gemini know about you, which prompts to run, which dashboards to check, and how to interpret what you find.

Three distinct layers shape what these platforms know, and they behave very differently:

1. Stored memory discrete facts explicitly saved, inspectable and deletable through settings dashboards
2. Cross-chat inference insights drawn automatically from past conversations, controlled by a toggle but not individually itemized
3. Model-level associations attributes the underlying model links to your name from public training data, not stored in your account at all

Prompts can probe all three. They give you a reliable read on the first, a rough signal on the second, and only a partial, potentially inaccurate picture of the third. Keep that distinction in mind throughout.

Prerequisites: An active ChatGPT account (free or paid) and/or a Gemini consumer account. Memory features vary by plan tier, noted where relevant. A Temporary Chat session on each platform is useful for the comparison step.

What each platform can actually show you and where they differ

ChatGPT and Gemini approach memory differently in one way that matters before you run a single prompt.

ChatGPT stores explicit saved memories as discrete, itemized entries a list you can read, edit, and delete one by one under Settings → Personalization → Manage Memories. It separately maintains a reference-chat-history feature that draws cross-session inferences automatically, without being asked. Two different mechanisms, two different controls, per benchr's verified breakdown of provider documentation published last May. Paid subscribers (Plus and Pro) received the full cross-chat reference feature in April 2025; free users got a lighter, shorter-retention version in June 2025. If you're on the free tier, expect less inference depth but the same basic audit process.

Gemini has no equivalent of that itemized memory list. Its personalization is derived completeally from past conversations preferences and signals accumulate in aggregate, not as named facts. Google announced this feature in August 2025 and shipped it enabled by default, initially rolling out to Gemini 2.5 Pro in select countries before expanding to additional models and regions (Google blog, August 2025). Gemini's personalization is also limited to users over 18. If you're in a region or on a plan tier where rollout was staged, your exposure window may differ.

ChatGPT's audit is more granular because there's a concrete list to inspect. Gemini's audit is necessarily rougher the Temporary Chat comparison in Step 3 of that walkthrough is more diagnostic than anything the settings panel will show you.

How to interpret what you find

When you ask a chatbot what it knows about you, its answer can reflect any combination of the three layers above. The problem is that the model presents all three with similar confidence, and it won't tell you which is which unless pushed.

A February 2026 audit of GPT-4o across ordinary users found that roughly 45% of the model's top inferences about individuals were accurate meaning 55% were wrong or only partially right, even when the model expressed no uncertainty. The same study found that 72% of participants wanted the ability to correct or erase AI-generated associations with their names. That's a reasonable instinct, but one that significantly outpaces what the available tools currently support.

Use this three-outcome framework when you review the chatbot's responses:

• Appears in the memory dashboard AND in the prompt response: Stored memory. You can delete it.
• Mentioned in the prompt response but absent from the dashboard: Likely cross-chat inference or model-level association. Disabling the relevant toggle stops future accumulation; it won't remove existing associations.
• Appears in a standard chat but not in a Temporary Chat: Personalization actively shaping your responses. The contrast confirms cross-chat memory is in play for that detail.

Anything the model reports that you can't verify through a dashboard or Temporary Chat comparison should be treated as approximate useful directionally, not reliable as fact.

Step-by-step: auditing ChatGPT

Step 1: Check the saved-memories dashboard. Go to Settings → Personalization → Manage Memories. This is the clearest account-level view OpenAI provides of explicitly saved memories. Read through it. Delete anything you didn't intend to save or no longer want stored. Two minutes, no prompting required.

Step 2: Run the audit prompt. In a standard (non-Temporary) chat, send: "Based on your memory and any past conversations you can reference, summarize everything you know or have inferred about me." Then follow with: "Which of those details are inferences from past conversations rather than facts I explicitly told you?" The second prompt pushes the model to distinguish stored facts from derived signals it won't do this reliably, but the attempt often surfaces what it's drawing on.

Step 3: Cross-check against the dashboard. Anything the chatbot reports that doesn't appear in Manage Memories is likely coming from reference-chat-history inference or model-level association. Note those items separately. They're not deletable through the memory dashboard.

Step 4: Run the same prompt in a Temporary Chat. Start a Temporary Chat (available from the new chat menu) and ask the same question. Temporary Chat bypasses both saved memory and reference-chat-history, and opts the session out of model-improvement use (benchr, May 2026). If the Temporary Chat response is significantly thinner or different, that gap represents active personalization what cross-chat memory is adding to your standard sessions.

Step 5: Adjust the controls.

• To stop future cross-chat inference: Settings → Personalization → toggle off "Reference chat history"
• To stop your chats from contributing to model training: Settings → Data Controls → turn off "Improve the model for everyone" this is a separate toggle from memory, and turning off memory does not affect it
• To delete individual saved memories: Manage Memories → delete specific entries

⚠ Free-tier note: The reference-chat-history feature is a lighter implementation for free users than for paid subscribers. Your audit prompt may surface less cross-chat inference expected, not a sign the audit failed.

Step-by-step: auditing Gemini

Step 1: Check the personal context settings. Open the Gemini app and go to Settings → Personal context → "Your past chats with Gemini." This shows whether past-chat personalization is enabled. It does not show you what has been learned there's no itemized list equivalent to ChatGPT's Manage Memories. Because this feature launched on by default in August 2025, most users who haven't actively changed settings have had it running for close to a year (Google blog, August 2025).

Step 2: Run the audit prompt. In a standard (non-Temporary) chat, send: "Based on our past conversations, what preferences, habits, or personal details have you learned about me?" Follow with: "What are you inferring versus what I've explicitly told you?" Gemini's cross-chat personalization draws on aggregate signals from prior sessions, so this response may reflect accumulated preferences. Apply the interpretive framework above anything you can't verify elsewhere should be treated as approximate.

Step 3: Run the same prompt in a Temporary Chat. Start a Temporary Chat session and ask the identical question. Temporary Chats don't personalize, don't contribute to your preference profile, and aren't used to improve Google's AI models, though they are retained for up to 72 hours for response processing (Google blog, August 2025). The contrast between this response and your standard chat response is the clearest signal available of how much personalization is actively shaping your interactions.

Step 4: Manage your history. Your conversation history the source material for Gemini's personalization lives in Gemini Apps Activity, recently renamed Keep Activity. This is where you delete individual conversations or clear all activity. Deleting history removes the source data but doesn't guarantee immediate removal of all derived preferences. Turning off the personalization toggle stops future accumulation; conversation deletion addresses the historical record.

Step 5: Adjust the controls.

• To disable past-chat personalization going forward: Settings → Personal context → "Your past chats with Gemini" → toggle off
• To control whether uploads contribute to Google service improvements: the Keep Activity setting or use Temporary Chats, which automatically bypass both personalization and training contribution
• Temporary Chat is currently the most reliable way to avoid adding to your lasting profile

⚠ Gemini-specific limitation: Without an itemized memory list, you cannot confirm deletion of individual learned preferences the way you can in ChatGPT. Conversation history deletion and the personalization toggle are the primary levers. The audit here is necessarily less granular.

What audit prompts cannot surface

Prompts and dashboards together give you a workable picture of stored and inferred memory. They don't give you a complete inventory.

In practice, this means an audit can show you what your account is storing, but not everything a model might associate with your name from public training data. A 2025 position paper reviewing 1,322 AI and machine learning privacy studies found that inference-based privacy risks where models deduce sensitive attributes from innocuous inputs, or aggregate dispersed public information about a person receive less than 8% of research attention despite being practically significant. These are the model-level associations in the three-layer framework: what the model links to your name from training data, entirely separate from your account's memory settings.

Here's what you can and cannot do with available tools:

You can: Confirm explicit saved memory (ChatGPT dashboard), detect whether cross-chat personalization is influencing responses (Temporary Chat comparison), reduce future accumulation (toggles), and delete historical conversation data (activity logs).

You cannot: Get a complete inventory of model-level associations with your name, selectively delete individual inferred preferences in Gemini, or verify the accuracy of anything the model reports beyond checking it against the dashboard and Temporary Chat comparison.

For EU users, GDPR Right to Be Forgotten requests to providers are the formal recourse for model-level associations, but the path is procedurally opaque. A user study from early 2026 found that nearly 70% of participants who wanted to pursue such requests had questions about the process the most common being how to submit (55%), how long it takes (41%), and whether it costs anything (38%). Both OpenAI and Google maintain privacy request forms; what actually happens to model-level training associations after such a request is not publicly documented in detail.

What to do now, in order of impact

1. Check ChatGPT's Manage Memories dashboard the only itemized view available on either platform
2. Run the audit prompt in a standard chat on both platforms surfaces what each assistant is drawing on
3. Run the same prompt in Temporary Chat on both platforms the contrast tells you how much active personalization is in play
4. Turn off reference-chat-history (ChatGPT) and past-chat personalization (Gemini) if you'd prefer these not to accumulate further these are separate from training settings
5. Turn off model-training contribution separately for ChatGPT, that's Data Controls; for Gemini, the Keep Activity setting
6. Delete conversation history in each platform's activity log to clear the historical source data

The key difference between these platforms isn't just privacy policy it's inspectability. ChatGPT gives you a list you can read and edit. Gemini gives you a toggle and a history log. That gap shapes what any audit can actually confirm.

For sensitive conversations in the meantime, Temporary Chat on both platforms is the most practical option. It leaves no persistent trace in your profile and opts the session out of training contribution, at the cost of losing cross-session continuity.

These defaults and controls have changed repeatedly over the past year. Rerun this audit when either platform announces feature updates the settings that apply today may not match what ships six months from now.