Gemini has a Memory Feature Too

The recent hype and ecstatic cheers about OpenAI's new memory feature created a swarm of happy users and some trepidation about security and privacy.

One of the most significant issues with ChatGPT is fundamental to its station in life. It is a SaaS application that serves as an outpost for generative AI work. It’s a silo of activity. It has no relationship with your data unless you inject it through a plugin or other means of copying all that you know; perhaps much of what you’ve forgotten and, at times, still matters.

GPT Memory is the latest attempt to make it seem like your investment in learning and using ChatGPT has a higher ceiling. A persistent memory from prompt to prompt is certainly beneficial. It creates continuity for prompts that are similar.

To access this feature, you must agree to invite OpenAI to use the data you place in memory for future model training. If you opt out and use the temporary memory feature, they won’t use that data for training, but we don’t need an AI to calculate the probability they will harvest this data to profile you.

This gating approach is tantamount to hosting a contest with the precondition of buying something, a practice that was outlawed in the US long ago. This post is not about privacy and economics, so we’ll put a pin in the deeper nuances of data risk.

Setting this obvious data-grab aside, ChatGPT Memory is a wonderful addition. It will accelerate productivity and add simplicity for many generative AI power users.

It’s a winning feature. How you get there is a bit sketchy.

Let’s assume you leave the default-in setting enabled for this new feature. And you decide to enter something innocuous as your location so that your generative AI queries can factor this information into future conversations. Maybe you store your profession and even some achievements for generating marketing content.

This data is now technically owned by OpenAI to do as it pleases.

There is no such persistent memory equivalent in Gemini. Or is there?

Your Google account is typically used to access Gemini, and indeed, this is the case for paid accounts. Gemini also has visibility into your Google Drive, Google Documents, and PDF files stored in Drive.

Generative AI Memory is Baked into Gemini

For Gemini-Advanced and Pro, you can create and manage many memory classes by creating instances of persistent information useful in prompts. You achieve this without the need for a new feature. It requires only the skills you already know well - Google Docs.

Step 1: Create a document; call it “My Credentials”.

Step 2: Add your credentials to the document. Be brief but clear.

Step 3: Save it as a PDF(1) and upload (2) it into Google Drive.

Step 4: Create a Gemini prompt to use your new memory document.

Step 5: Ask a question that only the memory would know.

Here’s an example:

“Name some of the companies mentioned in the credentials.”

Recalling the “Memory”

In the first prompt (shown here), I have established the My Credentials.pdf file as relevant to the conversation, and Gemini has confirmed this request. That’s all you need to do.

Please make sure to carefully note the prompt to recall memory. I did not need to specify where the document was. Gemini provides popups for general references to Google Docs and Drive features, but they are unnecessary when referring to a discrete document title, as I have done in the above example.

References made with @… feature extensions allow you to address collections of artifacts. Example...

List all @Google Docs that mention “CyberLandr” in a table with title, synopsis, and document size.

If you use a prompt like this, Gemini gets confused. Best to just reference the document name (I’ve learned the hard way).

Generate a list of experiences from @Google Docs [My Credentials.pdf].

I think this is a bug, or it literally assumes the extension's plurality.

Next, we test the memory. The result is spot-on. No drift. No hallucinations.

This demonstrates how to create memories for Gemini and works equally well in Gemini-Pro.

But There’s More…

Imagine a collection of “memory” documents for many generative AI tasks. You are not bound by any limitations to create broad collections of specialty documents for memorization. Furthermore, you can organize these in folders; Gemini doesn’t care where these documents are stored unless you give it explicit instructions.

Please use all PDF files in @Google Drive folder [Project A] for this conversation.

Wait! What?

You see where this is going, right?

Share

(1) I’ve noticed that PDFs tend to work better than Google Docs. Mileage may vary.

(2) You needn’t be concerned about where it lives. But give the GOOG a few minutes for it to save and settle in before inferencing with it.