Do ChatGPT, Gemini and Claude have a database?

A quick question

You type a question into ChatGPT, Gemini or Claude, and the answer comes back in about 2 seconds. Where did that answer come from? Did the tool look it up in a database?

Is it a search?

Most students say yes, because it feels like a search. You ask, and the information comes back.

But think about how a database search really works. If the data is there, the matching rows come out. If it is not there, you get "not found". A database never makes up a row.

These tools rarely say they don't know. They answer confidently, even when they are wrong. So the answer is not really coming from a database, and something else is going on. Let's try to understand what exactly is happening.

Think of an exam

Say you appeared for an exam. You read a question on the paper and write the answer. Your textbook is shut, so you are writing from your preparation.

How did you know the answer? You studied the topics, and you solved practice papers again and again. That practice is what lets you answer by yourself.

Now say a question comes up that you never prepared for. You won't leave it blank. You will write something, and it may be right or it may be wrong.

The engine learns the same way

The engine inside these tools learns like a student with practice papers. It reads a huge amount of text, again and again. Each time, it sees some words and guesses the word that comes next.

The engine plays this guessing game billions of times. After each guess it adjusts a little, and slowly it learns the patterns of the language.

What it means: our toy keeps a small table of which word came after which, so it can honestly say "no idea". A real engine keeps no such table, only numbers. So it always writes something, like you with the exam question you never prepared for.

What happens when you ask

When you ask a question, your text is first turned into numbers. Those numbers go into the model, which is itself a huge collection of numbers. Some maths runs on them, and out comes another set of numbers.

That result is turned back into text, one word at a time. Each word is the model's best guess for what comes next.

Where would a database even fit?

So what does the model keep from all that reading? It keeps no text, only numbers: billions of them, tuned to the patterns in the text. Could those numbers hold a full copy of everything it read? Let's check with real numbers.

Meta trained its open model, Llama 3, on over 15 trillion tokens. A token is a piece of a word, and one token is about 4 characters of English. The rest is our own arithmetic:

60 TB of text cannot fit into 16 GB, a space about 3,750 times smaller. So the training text is not copied into the model. What the model keeps is the patterns in that text.

There is one exception. Think of a song in a language you don't know. You can still sing it word for word, without knowing what it means. The engine is similar with text it saw very often: that text can come back exactly.

So there is no database inside the engine, but it is not a blank slate either. OpenAI says the same in plain words: its models do not store copies of the data they were trained on.

The engine has a name: LLM

This predictive engine inside the tools has a name. It is called a large language model, or an LLM.

An LLM holds everything it learned as numbers, which are called its parameters. So that kind of memory is called parametric memory.

The app around the engine

So is ChatGPT an LLM? Not exactly. ChatGPT is the app you talk to, and the LLM is the engine inside it.

Here is the important part: the engine does not remember your chat. Every time you ask, the app sends the whole conversation so far, along with your new question. The engine reads all of it, and answers.

So who does the remembering? The app does. It saves your chats and keeps a Memory of what you shared.

The app can also search the web and read a file you upload. Then it packs the right pieces in with your question and hands them to the engine.

Engineers call this arranging around the engine orchestration. The engine itself still has no database.

The price of guessing

Is guessing from patterns a bad thing? No. Because the engine works with patterns instead of stored pages, it can explain an idea in fresh words, not just fetch a saved one. But that ability comes with two costs.

First, it can be confidently wrong, like you with the exam question you never prepared for. OpenAI's own researchers say these models are rewarded for guessing, instead of saying "I don't know".

Second, its learning stops on a date. To test this, I used Gemini 3.8 Flash in Google AI Studio, with Google Search switched off. I asked: has the movie Avatar: Fire and Ash been released?

It said no, and that the film was due on 19 December 2025. In fact, the film came out on that date, and it passed $1 billion by 4 January 2026. The engine simply never read about it.

So how can we help it? Think back to the exam. Most papers have a comprehension section: a short passage, then questions on it. You answer from the passage, not from memory.

I gave Gemini the same kind of paper: a short passage about the movie, an empty line, and then my question. Copy it and try it yourself, in any AI app.

This time it said yes, because the passage makes the date clear. Now read the note under that answer. From its parametric memory, the model still says the film has not been released. The passage gives the right answer, while its old memory argues back.

Putting the information into the question like this is called context stuffing. It is exactly what these apps do with your chats, your files and the web.

The takeaway

1. ChatGPT, Gemini and Claude are apps. The engine inside each one is an LLM.
2. The engine has no database. It keeps patterns as numbers, and predicts the next word.
3. The app keeps your chats, your Memory and your files, searches the web, and hands the right pieces to the engine.

That leaves you with a question. If the engine has no database, how would you give it your own data, such as your college notes or your shop's price list?

Related