What Does an AI Model Actually Remember?
Understand that a trained model keeps patterns, not its training data, and why that makes its memory blurry.
Here’s a fact that surprises people: once training is done, the training data can be thrown away. The model doesn’t keep a copy. All that’s left is the weights.
Remember the km-to-miles model? It learned from a handful of examples, and what it kept was a single number: 0.62. Ask it about 7 km and it doesn’t look anything up. It multiplies.
That makes a model’s memory a lot like yours. Things it saw thousands of times stick, sometimes word for word: the capital of France, the opening line of a famous novel. Things it saw once blur or vanish. And when asked about something blurry, it doesn’t say “I don’t remember”. It produces whatever the patterns suggest, fluent and confident. That’s one big root of the made-up answers called hallucinations.
One more thing: it doesn’t learn from your chats either. While you use it, the weights are frozen.
Key takeaways
- After training, only the weights remain. The training data isn’t stored inside the model.
- Weights hold patterns. Ask about one specific example and you get the pattern’s estimate, not a recollection.
- Things seen many times can be recalled almost word for word; rare things blur.
- A blurry memory still produces a confident answer, which is one root of hallucinations.
Check your understanding
Go deeper (more technical)
The numbers: Llama 3 8B has about 8 billion parameters at 2 bytes each (16-bit numbers), so 16 GB. Its training text, about 15 trillion tokens at roughly 4 bytes per token, is around 60 TB. That’s a few thousand to one.
Models do memorise some text verbatim, especially passages repeated many times across the web. In 2021, Carlini and colleagues extracted exact training snippets from GPT-2, including names and phone numbers. Duplicates in the training data make this much more likely: one more reason to clean data.
Chat apps that “remember” you aren’t changing weights. They save notes about you in an ordinary database and quietly add them to the text the model reads. Track 2 explains how that text, the context, works.