It is like “10 LIST” quine.
For the machine code, the memory is THE source, and this code simply reads it because it can.
Imagine a quine in machine code running on an architecture where it is not allowed to read the code segment. This quine will not work there.
The REAL quine should generate the output purely via output data transformation encoded INTO the source, without any assumption of the runtime.
Those strings might or might not be kept in a different segment, depending on the architecture, but can you define "machine code" in an abstract enough way to allow some kind of string printing in general without self-reference?
Things are a bit more murky with variable-length instructions, but basically you are not allowed to read as data any address that is ever read as part of instruction decoding.
How do you know which addresses those will be? If the underlying architecture and OS permit reads or writes of code segments and/or execution of data segments, this set of addresses is literally undecidable.
You don't have to know these adresses beforehand, so you can just run the program to collect them and check. Or you roll up your sleeves and do some maths to compute (an overapproximation of) that set for the specific program you're interested in, which may be your only option if your architecture is nondeterministic.