hi gesteves!
I read your question, and I think it is readily seen to be not markov (because it is easily seen that P(X_{3(m-1)+3}=1|X_{3(m-1)+2}=1 \mbox{ and }X_{3(m-1)+1}=1) does not equal P(X_{3(m-1)+3}=1|X_{3(m-1)+2}=1)). In other words, since X_{3(m-1)+3}, X_{3(m-1)+2}, X_{3(m-1)+1} are giving information about the *same draw* from the mth box, most probably these variables are not independent and the proof should take into account of this.
Also note that X_{3(n-1)+i}, X_{3(m-1)+j}, 1\leq i, j\leq 3 are independent when m and n are different, as they correspond to different draws.
Since X_{3(n-1)+i}, X_{3(m-1)+j}, 1\leq i, j\leq 3 are independent when m and n are different, then P(X_{3(n-1)+i}=l|X_{3(m-1)+j}=k) = P(X_{3(n-1)+i}=l) = \frac{1}{2}, 0\leq l,k\leq 1 for different m and n.
As to the case when m and n are the same, it is necessary to calculate the probability explicity. But amazingly you will find that P(X_{3(m-1)+i}=l|X_{3(m-1)+j}=k)=\frac{1}{2}, 1\leq i,j\leq 3, 0\leq l,k \leq 1. For example, P(X_{3(1-1)+2}=1|X_{3(m-1)+1}=1)=P(\mbox{1 or 2 in the first draw}|\mbox{1 or 4 in the first draw}) = \frac{1}{2}.
Since all conditional probabilities are essentially 1/2, I think the assertion thus holds.