You

**do not** average imputed datasets. What you work on is parameter estimates, never on the actual data. Like you fit your mediation model in dataset #1, then again in dataset #2, then again in dataset #3 where there are as many datasets as there were imputations. Once you do that you use

Rubin's rules to combine the parameter estimates and obtain their variance. Getting the point-estimates of the parameters is easy (you just take the average of regression coefficients, for instance). Getting the right variance (i.e. standard error) for those estimates is a little more complicated because you need to calculate both within-imputation variance and between-imputation variance.

If you were using another software (R or STATA) they do that for you. SPSS (or the PROCESS macro) cannot handle it.