world-knowledge recall

World-knowledge recall is the model answering from facts it absorbed during pretraining, without looking anything up. Ask the capital of France, who wrote a famous novel, or how photosynthesis works, and the answer comes straight from patterns baked into the weights when the model read a large slice of the internet and books. In this sense the model is a compressed, fuzzy memory of much of what humanity has written down, and a great deal of ordinary knowledge can be retrieved this way astonishingly fluently.

The catch is that this memory is lossy and frozen. Common, often-repeated facts are stored robustly; rare or obscure ones may be blurred, half-remembered, or confidently invented, which is one root of hallucination. And the knowledge is fixed at the training cut-off, so the model does not know what happened afterwards and cannot tell that it does not know. This is exactly why retrieval systems exist: when freshness or precise sourcing matters, you feed the facts in through the prompt rather than trusting the model's internal recall.