Gerome,
It might be possible to mechanically perform such an analysis with separate effects and their t-statistics as input (e.g., some software does allow this kind of analysis), but it would be incorrect because the correlation between the two tasks would not be properly handled.
To be able to use 3dMEMA, you would have to provide both the contrast of Task A and Task B and the associated t-statistic. This means that you have to run 3dDeconvolve and 3dREMLfit by concatenating the two runs of data.
Gang