Hi Heiner,
Reconstructing exact memory operands and their order for instructions in Google workload traces is not really feasible, you can only infer the presence of memory operands from the read/write records that follow an instruction.
This was part of the filtering process we used on these traces to protect Google's IP.
You might be able to get an approximation for this missing info by manually creating a canonical template for each category + operand pattern encountered in a trace and make up missing operands like immediates.
As a suggestion, instead of trying to strictly satisfy XED, you might want to consider bypassing your tool's decoding layer and populate whatever internal IR your tool uses for instructions with the info you have from the trace (assuming your tool can take invalid-ISA patterns).
-Enrico