You can find all information about XMP here: http://www.adobe.com/devnet/xmp.html
But the document on how to embed XMP in PDF refers to the PDF specification as authoritative. This specification states that the XMP metadata shall be embedded in a metadata stream and that the document metadata packet shall be referenced by a key called "Metadata" from the document catalog.
That being said, XMP was specifically designed in order to be able to be found and read (and sometimes updated) without understanding the file format it is embedded in; it contains a magic fingerprint sequence at its start for exactly that purpose.
(You'll find the PDF specification on the Adobe developer web site as well, even though the latest version is in fact now an ISO standard - ISO 32000 to be specific)