Google Gemini Gets a Memory

Key Takeaways

• Gemini has added a new memory feature where it will remember certain things about you.
• This will help the AI chatbot to produce personalized answers to specific queries.
• This feature is quite similar to ChatGPT's memory launched earlier this year.

Google has rolled out a new memory feature for its Gemini AI chatbot. You can now ask Gemini to remember your choices and preferences and it will generate outputs based on these memories.

This change was first spotted by TestingCatalog News on X. Later, Google released a 25-second video on X explaining the feature along with some use cases.

This feature is currently only available on the web version of the chatbot and only to Gemini Advanced users. However, it is soon expected to roll out to Android and iOS users too.

• Google sighted a use case example in the video, where it told Gemini to always give 2-3 subject line options while writing an email. Now, whenever the user enters a prompt to draft an email, Gemini will also return subject lines as requested.
• Another example is that you can ask Gemini to remember that you're a vegan. Now, whenever you ask for restaurant recommendations from the chatbot, it will only suggest vegan-friendly places.
• Or, you can tell Gemini that you're allergic to milk and it will only recommend recipes that do not have milk.

There's a dedicated Saved Info page' where you can view, edit, and delete all the saved information. Remember, Gemini will save this information until you delete it manually. A Google spokesperson also said that the saved information is not used to train the AI model.

Get more consistent, predictable results without repeating the same details, maintain control over your data, and experience Gemini that seemingly integrates into your workflow. - Google blog post

Also, every time Gemini uses a saved memory to produce an output, it will notify you about it. So, in case you want to delete or edit that specific memory, you can do so from the chatbot's Settings. Currently, the feature is only available in English.

Similar to ChatGPT's Memory

This new Gemini feature is very similar to ChatGPT's memory feature which was launched in September. With ChatGPT too, you can ask it to remember certain things and facts about you, which is reflected in the outputs.

Plus, just like Gemini, you can add and delete memories either from the conversation dashboard or through the Settings page.

While this may seem an exciting feature and a step towards the future, we've already seen how chatbot memories can be exploited.

Johann Rehberger, a security researcher, has already demonstrated how ChatGPT's memory can be tricked into believing incorrect details through prompt injection'. This is a method where LLM models start following untrusted sources for inputs such as blog posts or emails.

For example, ChatGPT could be tricked into believing that the user is 102 years old or the user is a flat earther. What's even more shocking is that prompt injection can be used to alter the long-term memory of the chatbot.

Although OpenAI has released a fix for this issue, the dangers of prompt injection still persist. Now, as Gemini has also adopted this feature, Google needs to be on the lookout for possible bad actors.

The post Google Gemini Gets a Memory appeared first on Techreport.