ChatGPT, Gemini, and Claude quietly get worse over time. This hidden setting fixes them

Affiliate links on Android Authority may earn us a commission. Learn more.

I’ve been using LLM-based AI in some capacity for roughly half a decade, and in that time, the industry has changed significantly. Not only have newer competitors like Gemini and Claude arrived on the scene, but older tools like ChatGPT have also evolved dramatically. We’ve seen better third-party app integration, increased reasoning capabilities, and countless other advancements. However, I would argue that the ability to retain memories matters most.

Retaining context from previous conversations helps the AI better anticipate your needs within a single session. It also allows the tool to learn more about you, ensuring that future conversations are tailored to your preferences. However, there is a real downside to memory accumulation that not everyone considers. I’ve found that starting with a new AI usually requires significant input to generate the responses I need and fully understand my user profile. Once it reaches this data “sweet spot,” the AI becomes much more pleasant to use. Yet, slowly over time, the experience often seems to degrade again.

Let’s discuss what causes this gradual erosion, along with some simple maintenance steps to prevent old AI memories from dragging down your experience.

Old memories can linger long after you no longer need them in a few different ways. Sometimes, random one-off questions or hypothetical tasks are the root of the problem. For example, years ago, I ran a mock simulation of how I would hypothetically invest in the stock market. At the time, I made it clear this was just for fun. Almost six months later, while planning real investments, I asked the AI a related question. It confidently referenced a fake stock I had previously “invested in” and informed me how much its value had increased.

This is why I recommend using your AI’s “incognito” mode for less important inquiries or casual experimentation. This simple step can prevent unneeded memories from forming in the first place. Of course, that is far from the only way unwanted data builds up.

Some of this is simply the natural drift of time. You may have given the AI instructions a year ago that were relevant at the time, but your work or personal goals have since evolved. Humans are not static, but an AI sometimes struggles to understand this evolution.

You might assume that newer instructions will naturally overwrite older, conflicting memories. For instance, your previous religious beliefs may have influenced how the AI interacted with you. If you later went through a period of deconstruction and shifted your beliefs, you might mention this in a new conversation.