Glancing through the source for XML::Dumper, it looks as if the memory_address attribute is used basically as a key to identify portions of the data structure which are references to the same entity. It then makes the reconstructed data point to the corresponding existing reference in the new copy. YAML (and YAML::Syck) provide a similar functionality with anchors and aliases.

So it doesn't look to be used verbatim, merely used as a unique ID for some portion of the serialized data (granted the name of the key is misleading and your confusion understandable). Just read the attribute name as id and treat it as opaque (as the module does) and don't worry about it.