Let's take a look at how the mainstream real-time voice-based AI tools measure up in a real-world use case.

The Primer

The most common interactions with AI (currently) involve text-based conversations. Your typical text back and forth. Like a chatbot. But voice-based AI is becoming more prevalent, and let's face it... it's where it's headed. So, I wanted to look at how the different "mainstream" platforms handle these interactions through a lense of a particular use case. I'll be comparing ONLY this particular angle/feature of these platforms. Not the entire platform or the accuracy of responses, etc.

For each of these platforms, I'll be referring to the experiences in a web browser rather than their particular mobile or desktop apps. Because that'd just be way too complicated :)

The Platforms in Question

• Claude
• ChatGPT
• Gemini
• Copilot

The Test Case

To start up an interaction with any of these voice agents in a browser and start talking to them. Share information which it will remember for later use. Career history, skillset, values, other interests, personal motivations, family member names, your pets, your area, etc.

The Reason

To inform an AI agent about yourself so that it can remember and use that information later for practical purposes. To help it make better decisions or provide more accurate responses about potential career moves, project strategies, personal shifts, etc.

The Prompt

Hey, [AI System] I'm thinking of doing X. Give me your thoughts based on what you know about me and what I've willingly shared with you.

The Question

How does that experience and prompt go for each of these voice agents?

Here are the results:

Claude (Anthropic)

When speaking to it, it'd reply, and during its reply it'd hear itself and get into an infinite loop of interrupting itself.

Unusable (0/10)

ChatGPT

Near perfect live voice chat experience. However, it tells you to let it know when you want to save information permanently. And when you do that, in some cases it will lie about whether it actually happened or not. If you tell it something like "Remember my brother's name", It will say it has saved that information permanently for the future. It has not. It's a privacy thing apparently built into the platform. But you only find this out after testing or pressing it to tell the truth.

Hidden long-term memory rules, but great when reminded of things (8/10)

Gemini (Google)

Different capabilities depending on the type of Google account you have. Gmail accounts have certain abilities. Google Business accounts have others. The lines are not clearly defined. After that, the long-term memory capabilies are confusing to determine, but in testing seem nearly non-existent. It doesn't remember almost anything.

Account dependent. Weak if any memory (3/10)

Copilot (Microsoft)

Pretty much the same live voice chat experience as ChatGPT. Extremely good. When you ask it to remember something permanently for the future, it actually does so. Without restrictions. If it says it'll remember, it does. Always. The only problem is the configuration abilities when it comes to controlling the tone of the agent and giving it global guidance before speaking with it. ChatGPT is far better at that. Also, the user interface is clunky. You have to switch between voice and text modes too often instead of things staying in the same area.

Weak configuration and UI, but only one to have actual persistent/useful memory (9/10)

Conclusion:

All of these platforms are working on this situation from different angles and through different prioritization lenses. Some are better in certain aspects, while others excel in others. There are also dozens of other platforms that I did not cover here which may tackle this use case perfectly. They just don't have the data/breadth to make it so this