After sitting through nearly every AI session at the National Association of Realtors Legislative Meetings 2025 in DC last week (I still call it NAR Midyear), I was struck by how much confusion still exists. Too often, the AI "experts" on stage gave incomplete, misleading, or simply wrong information. 

That means agents heard things about AI and what AI could do that were untrue. Unfortunately, while we can blame AI for hallucinating and getting things wrong, those educating agents must ensure what they share is current and correct. It’s encouraging to see AI front and center, but let’s not make assumptions. 

Admittedly, AI is rapidly changing, so even this column could be outdated next week. But that makes it even more crucial that the AI information shared with real estate pros is updated and verified while avoiding generalizations. 

The memory myth 

The most common blunder was this: assuming all the top chatbots have memory. More than one speaker said you could teach “your favorite chatbot” to write like you, speak in your tone, or follow a master prompt across sessions. 

That’s just not true. 

First, only paid versions of chatbots have persistent memory. 

And even then, not all paid versions are created equal. Claude Sonnet 4, Perplexity Pro, and Grok may be fast and flashy, but in regular chat mode, they forget everything once the session ends. That’s a problem.

For real estate agents, it means you’d have to re-teach the AI your voice, style, and preferred responses every single time. You also can’t teach them different voices and writing styles. They certainly won’t recall a master prompt across sessions. 

It’s another reason using free versions is a bad choice: without persistent memory, you're starting from scratch with each prompt. If you want an assistant who learns how you work, you need one who remembers. 

Right now, real memory is still mostly a ChatGPT thing. You'll find it in ChatGPT Plus and Pro, but it's especially powerful in ChatGPT Teams, which we use. 

Outside of that? Only Gemini Advanced and NotebookLM come close, with comparable persistent memory, though each comes with its own quirks and caveats. 

Persistent memory matters because it saves time, boosts consistency, and helps you scale your efforts. 

Imagine not having to repeat your tone, writing preferences, or language when creating a Facebook post, blog, or listing presentation. It's the difference between a productivity drain and efficiency gain. 

File size stoppers 

Here’s something no one on stage talked about: file upload limits. Most of the AI speakers at NAR admitted they used one of the paid versions of ChatGPT. However, the specs and capabilities for ChatGPT are not the same for other bots. 

For example, if you're planning to upload PDFs, listing presentations, inspection reports, or CMAs to your chatbot for background for your prompt, think again. Some bots choke on even modest file sizes. 

For example, the paid version of Claude has a 30 MB limit and PDFs no more than 100 pages. Perplexity? 50 MB. Grok? Same ballpark. Gemini just raised its cap to 100 MB. Even NotebookLM at 200 MB is half the limit of ChatGPT. 

In contrast, ChatGPT (Pro and Teams) handles up to 512 MB per file. That means you can drop in a dozen different files and get valuable insights without trimming, compressing, or splitting your files first – or having to create multiple sessions. 

If you're using AI to summarize MLS data and dissect research reports, you need a chatbot that can handle big uploads without giving you a file-too-large error. 

Not disclosing that there are limits will end up frustrating a lot of agents and could turn some off from using AI if they keep being told it does one thing, yet it does something else. 

To avoid AI tools that limit you, research your chatbot's memory and upload capacity before you commit. (-Kevin)