Knowledge lives in the model's parameters, but as distributed representations spread across many weights and layers, not as explicit records. You can't open the model and find a row that says "capital of India = New Delhi" — the same fact may be encoded across many weights simultaneously.

The model learns patterns from training data and encodes them as high-dimensional representations rather than key-value entries. This is exactly what enables generalization — the model learns relationships, concepts, grammar, style, and reasoning patterns, not just isolated facts — but it's also why inspecting, editing, or deleting one specific fact is hard: correcting "Person X works at Company Y" may require changing weights that also encode many unrelated things, with no guarantee the edit won't have side effects elsewhere.