Large language models like those powering ChatGPT are stateless, meaning they retain nothing between conversations. What feels like memory is actually a separate system of five distinct layers bolted onto the model, each operating independently and each something you can inspect, edit, or delete. This distinction matters far more than a technical footnote, because it means you have genuine control over what your AI assistant stores about you, once you know where to look.

Does Your AI Assistant Actually Remember You Between Conversations?

The short answer is no. When you return to ChatGPT, Gemini, or Claude after a break, the model itself has no built-in recollection of your previous chats. A trained model is a fixed set of mathematical weights that does not change when you interact with it. It reads your prompt, generates an output, and holds nothing over to the next request.

What happens instead is this: a memory system sits around the stateless model. When you return, that system retrieves relevant details from past conversations or a saved profile and re-injects them into your prompt as context, just before the model runs. The model then answers as if it remembered, when it was actually handed the details a fraction of a second earlier. One Reddit commenter captured this distinction sharply when ChatGPT's memory upgrade landed, noting that the system "built in a search feature. It just searches stuff linked to your account."

One Reddit

What Are the Five Layers That Create the Feeling of Memory?

Because memory lives in an external layer instead of inside the model itself, you have more control than you might assume. You cannot rewrite the model's weights, but you can view, edit, and delete the stored content the system feeds back into it. These five mechanisms get lumped together as "memory," yet each lives in a different place, lasts a different length of time, and has its own deletion path.

• Context Window: Everything in your current chat session, which ends automatically when the conversation closes.
• Referenced Chat History: Automatic retrieval from past chats stored in your account, controlled through history and memory settings.
• Saved Memory: Facts you state explicitly or the system infers about you, stored in a separate profile store attached to your account.
• Document Retrieval (RAG): Chunked and embedded files that the model searches when your prompt matches stored content.
• Model Training: The only layer that actually changes the model itself, occurring when vendors use your data to improve future versions.

Only one of these five layers, model training, alters the model's weights. The other four simply re-feed text back into the prompt. This is the load-bearing idea for understanding how much control you actually have.

How Do Default Memory Settings Differ Across AI Assistants?

The philosophies split sharply among the major assistants. Three of the four lean toward remembering by default; one does not. OpenAI states that ChatGPT "now references your recent conversations" as of an April 2025 update. Google's Gemini "learns from your past chats to understand more about you," according to its Saved Info settings page. Microsoft's Copilot includes memory and personalization features by default.

Anthropic takes the opposite stance. Claude does not retain conversation content by default, a choice that reflects a fundamentally different philosophy about user privacy and data handling. Users running both tools side by side describe this split vividly. One person on Reddit noted that ChatGPT uses a unified memory model where information shared in one conversation surfaces in others, sometimes unexpectedly. By contrast, Claude silos memory by project, so it knows nothing outside a specific conversation thread. The practical lesson is clear: do not generalize from one tool. Check where your assistant sits, by vendor and by tier, before assuming it remembers everything or nothing.

What Happens When You Delete a Chat?

Here lies the single most consequential misconception in the entire topic. If you delete a chat, the AI usually does not forget everything it saved about you. Saved memories live in a separate storage layer from your chat history, so removing a conversation does not remove the facts the system saved. OpenAI confirms these are independent layers that must be deleted separately.

This confusion is understandable. In a Pew Research Center survey, 81% of Americans who had heard of AI believed collected information would be used in ways people are not comfortable with. The system's design makes this worry rational. A memory you created months ago can resurface in a brand-new conversation long after the original chat is gone, because it lives in a dedicated profile store attached to your account, not in the chat history itself.

How to Stop an AI from Remembering You Across Sessions

If you want to prevent your AI assistant from building a profile about you, work through these four steps, treating each one as separate. Switching off one layer does not touch the others.

• View Saved Memories: Open your assistant's memory or personalization settings, where saved items are listed individually so you can see exactly what the system has stored.
• Toggle Off or Delete: Remove memories one by one, or disable memory entirely so no new ones are created going forward.
• Opt Out of Training: In data controls, turn off model training. On ChatGPT's Free, Plus, and Pro tiers this is opt-out, not opt-in, meaning you must actively disable it.
• Use Temporary Mode: Temporary chats do not save to history or create memories, but in ChatGPT the mode must be chosen before the chat starts and cannot be switched mid-conversation.

One important caveat: removal lags behind the delete button. OpenAI schedules deleted chats for permanent removal within 30 days, so the data does not vanish instantly. If you are concerned about privacy, plan accordingly and do not assume deletion is immediate.

The broader implication is reassuring. Because memory lives in an external layer instead of inside the model, you have more control than you might assume. You cannot rewrite the model's weights, but you can view, edit, and delete the stored content the system feeds back to it. Understanding this five-layer architecture transforms memory from a black box into something transparent and manageable.