I told an AI assistant once that I’d take the window seat on a flight with my wife. One time. Weeks later, it was still booking me a window seat every time, without me asking, because somewhere along the way that offhand comment turned into a rule it decided to keep.
That’s a harmless example. But the same mechanism is now sitting inside the AI tools accountants and finance teams use every day, and it’s worth understanding before it costs you something real.
Memory is now turned on by default across Claude, ChatGPT, and Gemini. Each of these companies has landed, more or less, on the same conclusion: users like it when the assistant remembers things. It feels like the tool is learning you. It isn’t, technically. You’re not training a large language model just by chatting with it. What’s actually happening is simpler, and a little more concerning: the tool is quietly updating a memory file behind the scenes, and it decides on its own what’s worth saving.
That’s the part that should give accountants pause. Claude, ChatGPT, and Gemini all manage this themselves. You can make one comment in one chat, and the model can decide that’s a “core memory” worth keeping and it may never tell you it did. If you never go looking, you’d have no idea it remembered anything at all.
Sometimes that’s genuinely useful, it remembers you like flying Delta, or that you want variance reports formatted a certain way. But sometimes it’s a problem. Tomasz Tunguz wrote about this in a blog post that came out just a few days ago, about how long an AI agent’s “life” should be. His example: tell your assistant in March that you have a cold and to cancel your morning meetings. By November, it might still be avoiding morning slots because that one-time comment became a standing rule, and nobody ever went back to unwind it.
Now translate that into the world of accounting and finance. Picture telling your AI tool, in the middle of a close, “just exclude that vendor variance this month, it’s a known timing issue.” That’s a completely reasonable thing to say once. The risk is that the model treats it the same way it treated the window seat request, as a rule to keep applying, without telling you. Next month, it might just decide the variance is “already explained” and move on, and you may never see the reasoning it used to get there, because these tools are giving us less of their raw logic and more of a summary these days.
That’s why I think of this as a new internal control question, not just a tooling quirk. If you’d never let a manual journal entry override sit in the system without a documented reason, an owner, and an expiration date, you shouldn’t let an offhand comment to your AI tool sit there indefinitely either.
The tricky part is that none of these tools is transparent about this by default. It’s a bit of a black box until you actually go look. The good news is that all three of the major providers give you a way to look. In Claude, it’s under Settings, then Memory, it saves “topics” in real time, and it keeps a separate memory per project. ChatGPT has a similar setting plus a “reference chat history” toggle you can turn off separately. Gemini calls it Personal Intelligence, and it can pull from your connected Google apps as well as your chat history. All of them will tell you what’s turned on if you just ask.
So here’s what I’d actually do this week: open a new chat with whatever AI tool you use for client or firm work, and ask it directly: What do you remember about me and my work? Then go check the memory settings themselves, and check whether “reference past chats” or “reference projects” is turned on too, since that’s a second, separate way old context can sneak back in. Edit or delete anything that’s stale. If your firm is on an enterprise or team plan, don’t assume memory is off. Check the admin console, because it varies by provider and it changes.
Key Takeaways
- Memory is on by default across Claude, ChatGPT, and Gemini, and each provider decides on its own what’s worth remembering — often without telling you.
- A one-time comment or override to your AI tool can become a standing, undisclosed rule, the same way a comment about cancelling morning meetings can still be shaping your calendar months later.
- Treat a “just exclude this one time” comment like a manual journal entry override — it needs a reason, an owner, and an expiration.
- Run a memory audit: ask a new chat what it remembers about you, check the memory setting, check the “reference past chats” setting, and prune anything stale.
- Don’t assume your firm’s enterprise plan has memory off — check the admin console, since policies differ by provider and change over time.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply