Somehow, a normal search spiraled into a full Black Mirror moment. I asked Gemini to recommend a place for a date night, and it suggested an Italian bistro.

Then it said something that caught me off guard. Gemini recommended the truffle risotto and casually mentioned that I’d been there with Sarah last spring and enjoyed it.

There’s just one problem. Sarah and I broke up six months ago, and it didn’t end well.

The risotto is tied to a memory I’d rather forget, and I absolutely don’t want my AI assistant bringing it up.

I want Gemini to forget all of it — Sarah, the risotto, all of it — and reset.

But when I tried to do that, I realized that while Gemini is good at remembering, it can’t let go easily.

Gmail logo surrounded by ghosts and Gemini icons against a dark background.

Why Gemini remembers things you thought it would forget

A man’s head opened, showing his brain powered by Gemini AI
Credit: Lucas Gouveia / Android Police | Prostock-studio / Shutterstock

To understand why I couldn’t “Men in Black” Sarah out of Gemini’s memory, we need to look under the hood.

In the AI marketing talk, the word “memory” carries a lot of weight. For a large language model like Gemini, memory is two buckets of data.

Active memory and what you explicitly tell Gemini

The first bucket is what Google calls Saved Info or Instructions for Gemini. This is the explicit memory.

You might tell Gemini, “I prefer my code in Python without analogies,” or “Always assume I am writing for a senior management audience.”

This information is stored in Settings > Personal context > Your instructions for Gemini. It is transparent, easily editable, and works as a guideline for your conversations.

Passive context and what Gemini pulls from your account

The second bucket is the implicit memory, or the Personal Context. This is where Sarah lives.

Gemini has read-access to your Connected Apps —Gmail, Google Photos, Drive, Calendar, and more — across your Google account.

To remove implicit memories, go to myactivity.google.com/product/gemini, scroll through a chronological list of everything you’ve ever said to the AI, and delete the interaction.

If you talk to Gemini daily, this is a near-impossible task.

ChatGPT handles personal memory better than Gemini

Illustration of two checklists, one labeled Gemini and another ChatGPT, with several thumbs down around it
Credit: Lucas Gouveia / Android Police

Midway through my purge struggle, I decided to see if the grass was greener on the other side. I logged in to ChatGPT.

I went to Settings > Personalization > Manage Memories. I saw a bulleted, human-readable list of facts.

  • User lives in Sydney.
  • User prefers Python over JavaScript.
  • User is planning a two-week backpacking trip across Europe.

ChatGPT’s version was cleaner and even had an option to search for memories.

Since I don’t want to subscribe to GPT again as I’m happy with my Google AI Pro deal, I decided to use the Temporary Chat feature. It’s Incognito Mode for AI.

Google needs a forget button ASAP

A woman using a phone in bed, surrounded by floating Google Gemini and Workspace app icons-1
Credit: Lucas Gouveia / Android Police | yourphotopie / Shutterstock

If Google engineers are reading this (and let’s be honest, their AI is), I have a proposal.

We need a UI feature I’m calling the Flashy Thing Button, named after the neuralyzer from Men in Black.

Here is how it should work:

  1. Instead of deleting a chat, let me search for an [Entity] in Gemini Settings.
  2. The system should then show me the blast radius. For example, Gemini found 423 emails, 120 photos, and 50 chats related to [Entity].
  3. Give me a button to forget or at least block retrieval queries related to this [Entity].

If Gemini can filter out NSFW content or hate speech, it can filter out my ex.

Deleting AI memories is harder than it sounds

A brain connected to some laptops, the Gemini logo in the center, and some warning signs around.

I read some research papers on machine unlearning (yes, that exists). Apparently, teaching a neural network to forget is one of the hardest problems in computer science.

When a model learns something, it adjusts billions of weights (connections between neurons). Like a drop of dye mixed into a gallon of water, you can’t reach in and pick out the dye molecules.

To truly forget a fact that a model has trained on, you often have to retrain the whole model from scratch, minus that one piece of data. That costs millions of dollars and takes months.

However, a simple keyword block feature would still go a long way without any costly retraining.

The case for a right to forget in personal AI

Forgetting is a survival mechanism that helps us focus on the present through generalization and abstraction.

It’s certainly not a flaw. Forgetting is how we grow. How we forgive. How we move past trauma.

AI’s infinite memory runs against the very grain of what it means to be human.

This is one of my favorite quotes from the movie Eternal Sunshine of the Spotless Mind, quoting Friedrich Nietzsche:

Blessed are the forgetful, for they get the better even of their blunders.

Until Google gives us better AI memory-management tools, the safest way to use Gemini’s memory is not to use it, and even that comes with its own shenanigans.