Chatting Isn’t Training: Demystifying Memory in LLMs

Misconception: LLMs “hoover up” everything you type

Posted on October 15, 2025 in Blog Posts

Authors:

Nicole HennigChad MairnKevin Leary
An AI chatbot training on chats

A common fear is that anything typed into a large language model (LLM) is instantly absorbed into its training data, compromising privacy and intellectual property. This concern, while understandable, reflects a misunderstanding about how LLMs are trained and operate. Although chats may, if users don’t opt out, be anonymized and possibly used for training one day, training is not the same as chatting, so it helps to understand the difference.  

This post will demystify how training and chatting work, help users understand how LLMs handle their personal data, and allow users to make informed decisions about their privacy. 

The two parts of generative AI systems 

Generative AI systems are made up of two parts: a chat app and the underlying model, which powers the app. The chat app can keep a history of your previous chats, but the model only knows what it has been trained on, not everything you’ve asked it. 

Models are static files that are trained or updated infrequently. As a Claude engineer explains, “The model is stored in a static file and loaded, continuously, across tens of thousands of servers…. The model file never changes and is immutable once loaded.” Training a model like ChatGPT usually takes weeks to months, depending on the scale of computing power. It goes through several stages, and the whole life cycle (planning, experiments, training, alignment, evaluation) often spans 6–12 months before release. In short, your conversations are not automatically being used to train a model in real time. 

Learn more about training:

Chatbot history versus training  

The chat app keeps a history of your conversations for your own use so you can refer back to them and continue a conversation. If you give feedback that a particular answer is bad or harmful in some way—such as using the thumbs up or thumbs down symbol in ChatGPT— your conversation may undergo a safety review. In that case, the company’s authorized reviewers may examine the chats you tagged. Most model companies keep some interaction data for limited legal, safety, and quality purposes. Providers typically indicate that reviewed items are anonymized, but retention windows vary widely (e.g., from hours to months or longer in specific cases). OpenAI, for instance, retains chat data for 30 days unless there are other legal obligations.  

Data retention is not unique to AI models. Other digital services libraries use, such as discovery platforms and research databases, frequently keep interaction data and limited logs of IP addresses and time stamps to meet legal requirements, prevent abuse, and improve reliability. 

🔥 Stay up-to-date with LibTech Insights by signing up for our free newsletter. Just one weekly email with our new blog posts, top tech news stories, and other bonus content. Check out some posts from our archive:

How to opt out of using your conversations for future model training 

Other than safety reviews, what could your chat history be used for in the future? If you are using a consumer version of the product, it’s possible that the model company might use de-identified (that is, disconnected from your account) chats as part of a future training run. This gives the company examples of real-world use that can help models respond more effectively to user inputs and improve safety filters. 

If you would prefer not to have your data used in this way, you can opt out in the settings. All the major foundation models make it possible to opt out of training future models using your conversations. This applies to consumer versions of their products, whether free or paid. 

Education, enterprise accounts, and apps built on model APIs don’t train on your data 

Education and enterprise accounts are opted out of training by default to protect your institution or organization’s proprietary and sensitive information. Companies and universities pay for this privacy when they subscribe to these types of accounts: 

In addition, apps built on model APIs are not used for model training (see OpenAI, Anthropic, Azure OpenAI, Google Vertex AI policies, and Perplexity). This is also true of library vendor products that use APIs to power some features, such as Clarivate. 

Consumer accounts allow you to opt out 

If you have a consumer account, here is how to opt out of having your conversations used for future training. 

  • OpenAI, ChatGPT: Click your name, Settings, Data controls, and turn off the switch called “Improve the model for everyone.” (See Data Controls FAQ). 

A notable exception for consumer accounts is Google’s NotebookLM. Unlike many other tools, it does not train on your personal data, so there is no need to opt out (see: Learn how NotebookLM protects your data). Steven Johnson, Editorial Director at NotebookLM and Google Labs, explains that you can upload copyrighted materials and that “no material uploaded to the model is used to train NotebookLM or Google Gemini.” Johnson clarifies that if you have the rights to use copyrighted materials, you can upload them to NotebookLM. For each source, NotebookLM generates a “guide that identifies patterns from the information in the source” and then produces a summary paragraph, meaning you can safely upload licensed or copyrighted materials into NotebookLM as long as you don’t share the Notebook publicly. 

Temporary or “incognito” chats 

Another way to exclude your chats from future training data is to use the “temporary chat” feature (sometimes called “incognito mode”). This is helpful if you have not turned off the settings described above or just want to opt out of training for certain conversations. 

These links provide instructions on that feature. 

  • Copilot: Privacy FAQ. (Temporary chats not yet available, but some say it’s coming soon.) Note: You can also use Copilot consumer edition in a limited way without logging in. In that case, your conversations are anonymous by default. 
  • Perplexity: Account and settings. Click on Account, then choose Incognito. Note: You can also use Perplexity in a limited way without logging in. In that case, your conversations are anonymous by default. 

Memory and personalization  

Recently, some models have added memory features, which allow the app to refer to information you’ve discussed in previous chats and give answers that are more personalized to your situation. 

If you think about it, companies wouldn’t need to add such features if everything was continuously being “hoovered up” across your conversations. Models can’t remember across your conversations by default due to the “context window” (aka short-term memory). That is everything you type into the model, including any documents you upload, and everything it types back to you in one conversation. When you start a new conversation, it doesn’t remember what you talked about in previous conversations, unless you have activated a memory feature. 

It’s important to understand that different models have different sizes of context windows (i.e., how much can be remembered in one conversation). This is measured in “tokens” or parts of words. The context window size may change with new models, but generally, Gemini has the largest, followed by Claude and then ChatGPT.  

This will matter to you when you have a very long conversation. If your conversation (including any documents uploaded) goes past the size it can remember, it will “forget” information from the beginning of your conversation to make room for the new information. 

Memory features 

Many people find the new memory features in ChatGPT and other models useful. They allow the model to remember previous conversations so you don’t have to repeat yourself when discussing the same topic in a new chat. Good prompt engineering involves giving models plenty of specific context so they generate answers that are relevant to your task. In fact, context is so important that some propose that we use the phrase “context engineering” instead of “prompt engineering.” 

For example, you might want a model to remember your job role, the types of tasks you often do, and the types of audiences you usually work with. It can apply that information to future answers to make them more relevant. But if you sometimes use a model for demonstration purposes, you may want to exclude those conversations from its memory. 

Luckily, you can turn off memory in the settings or use a temporary chat for those conversations that you don’t want as part of the memory. 

Another method to use with any of the models that have memory is to ask in a chat, “What do you remember about me from past conversations?” From there, you can ask it to forget certain memories or add extra memories.  
 
Another new feature in many models is project-based memory. This allows you to create projects with a group of related chats and apply specific memory settings only for that group.  

Claude handles memory differently from other models. See “Claude Memory: A Different Philosophy.” 

A focus on better data for training 

As models aim to perform more complex “reasoning” tasks, companies are now training with better quality data (not just random web crawls). This requires curated datasets that simulate expert-level thinking.  

Because of this, they are moving away from the previous practice of using outsourced data-labeling firms that used low-paid labor from the Global South. Instead, they are using highly paid industry experts in fields such as computer coding, biology, and finance to help create more sophisticated, specialized data for model training. Here is an example recruitment message for this type of specialist, received by one of the authors of this post. Another of the authors of this post interviewed a colleague who does this work.  

“When I work on model training, my role is to help refine and evaluate outputs, so the AI learns from higher-quality feedback. This shift toward using domain expertise, rather than relying solely on massive amounts of scraped or outsourced data, helps create models that are more accurate, responsible and useful for real-world applications.“
 
– Samantha Stanich, Freelance Content Writer 

The importance of a nuanced understanding 

What are some dangers of believing the myth that everything you type is instantly used for model training?  

For one, it can lead people to avoid legitimate educational and assistive uses that are helpful for learning. For library workers, the myth flattens nuance, making it more difficult to ask vendors the right questions about privacy in their products. As information professionals, we care about protecting user privacy, so it’s important for us to understand these nuances while making things simple for our users.  

Here are some points to remember: 

  1. Enterprise and education versions of models don’t use your data for future model training. For example, see Enterprise data protection in Microsoft 365 Copilot and Microsoft 365 Copilot Chat.
  1. Consumer versions of models offer ways to opt out of future model training, and NotebookLM doesn’t use your data for training. Temporary chat is another way to keep your data private. See How can I protect my privacy when using ChatGPT & similar tools? 
  1. It’s a good idea to periodically check the settings in any model you use and make sure you’ve selected the options that matter to you. That’s where you’ll find out if anything has changed. 

We hope this has been helpful! We are working on some future posts about other misconceptions about generative AI. Feel free to contact the authors with suggestions or feedback. 

Subscribe to the LibTech Insights newsletter 🌟

Related Posts