Have you seen this ChatGPT warning before? Here’s a fix.

For the past several months, I've been keeping track of health issues in a ChatGPT chat. Things like dates, numbers, symptoms, and observations. I didn't set out to build a personal health record. It just evolved into one because it was the only method of tracking that I actually maintained over time.

One night I was waiting for it to finish answering and I read the little grey line under the message box for the first time. I'd scrolled right past it a hundred times. This time it landed, because the long conversation it was talking about was mine, and mine went back seven months with nothing saved anywhere else.

Then I started noticing _mistakes_

For the first few months, everything seemed fine.

Then, around month five, I started noticing details that didn't line up with what I'd recorded. I asked for a summary of the previous few months and got back a start date that was off by weeks, along with a number I knew I'd never entered.

I asked for a quick summary of the last few months and it handed me a start date that was off by weeks, sitting right next to a number I know I never gave it. Most of the response looked reasonable, but some of the details weren't coming from my notes.

It wasn't leaving a blank where the old entries used to be. It was filling the gap with something that sounded right and wasn't.

What surprised me wasn't that it forgot. Long conversations have limits, and I expected that. What surprised me was that it often returned an answer anyway.

That's when I realized I couldn't treat a chat as a reliable place to store long-term information. Once enough history accumulated, some details started drifting, and it became harder to tell what came from my records and what came from the model trying to fill in the gaps.

It wasn't just _one mistake_

At first I assumed it was a one-off mistake, so I kept testing it. I asked about things I already knew the answers to. Each time, it responded quickly and confidently, but some of the details were off. The more I checked, the more obvious it became that I couldn't rely on a single chat to preserve months of history accurately.

So I stopped. It was time to give the record somewhere real to live.

For ChatGPT, Empirical is just an app. You turn it on and it's there:

@installcard(https://chatgpt.com/apps/empirical/asdk_app_699db5f04b788191a4f9ee070d3e5d67, "Add Empirical to ChatGPT", "chatgpt.com · turn it on, no setup")

It also runs with Claude, and with Codex in the terminal, but ChatGPT was where my history already was, so that's where I started.

@installcard(/docs/install, "Install Empirical for coding agents", "Codex, Claude Code, and CLI workflows")

I told it to save _everything_

The first thing I did was move it all over. One message:

RECEIPT 02 · MOVING THE RECORDMEMORY STORED

I asked it to record all my data tracking in Empirical as separate memories. A few seconds later: Memory Stored. Every entry suddenly had a home outside the conversation.

Then I asked the question that started all of this. Given that warning about long chats forgetting things, does saving it into Empirical actually help, or is it just one more place for the same data to go stale?

Here is what it said.

RECEIPT 03 · MEMORY ENGAGEDMITIGATED

The answer: even if a future conversation does not contain all of this chat history, the health tracking can be retrieved from Empirical rather than relying solely on the conversation context. The forgetting was still possible. It just could not take the record down with it.

Beyond the _chat_

The point was never to make one chat remember more. It was to stop relying on a single chat to remember anything at all. My history now lives in Empirical, so I can start a new conversation and still have access to the dates, numbers, and notes I tracked months ago.

The record used to live in the chat. Now the chat is just one of the places I can reach it from.

So I'm not stuck in that one thread anymore. I can add an entry from ChatGPT on my phone right after an appointment, pull the last few months from Claude on my laptop, or have Codex check something in the terminal while I'm working.

Same records, whichever one I happen to have open.

Three things changed immediately:

1. It's not stuck in the chat. The chat is the part that forgets. Empirical holds the records off to the side, so a brand new conversation can pull up my real history instead of guessing at it.
2. It kept the real entries. Not a note that says I track my health. The actual dated entries, which is what you need when you want to answer "when did this start" half a year later.
3. I can reach it from anywhere. ChatGPT, Claude, or Codex. In the terminal on my laptop or in a browser on my phone. I can add a new entry or look an old one up, and it's the same memory either way.

The _takeaway_

What I took from it is simple. If you'd be upset to lose something, don't let it live only inside a chat. Health notes, the running log on anything ongoing, the stuff you keep meaning to put somewhere safer.

A chat is happy to forget it, and worse, happy to hand you a confident replacement.

I didn't quit using ChatGPT for this. I just stopped counting on the chat to be the thing that remembers. Empirical holds the record now, and I can get at it from whatever I'm using.

I don't think about that warning anymore.

All Empirical blog posts