The marginalized posterior of a parameter doesn't have to be symmetric and have its mode at the center, which would be the case most people think of (e.g., the Normal distribution) when envisioning the posterior. In fact, by choosing your prior appropriately, the effect you mention can happen even with a "nice" likelihood like that of the Normal distribution, but my example will be a little different.
Consider the posterior distribution for the mean parameter $\theta$ of an Exponential distribution with a prior distribution ($p$) uniform on $(0,100000)$: $p(\theta) = {1 \over 100000} \space \text{I}_{(0, 100000)}$. Let's assume the true parameter value $\theta = 1$, which renders the upper bound effectively meaningless when calculating the posterior, as it is far larger than any sample will ever indicate $\theta$ might be. The posterior distribution ($p'$) for a sample of size $N$ then becomes proportional to:
$$p'(\theta) \propto {1\over \theta^N}\text{e}^{-\sum{x}/\theta}\text{I}_{(0,10000)}$$
and, for all practical purposes, we can ignore the upper bound on the range of $\theta$. This is an Inverted Gamma distribution with shape parameter $N-1$ and scale parameter $\sum x$, and has a mean of ${\sum x \over N-2}$ but mode (MAP value) $\sum x \over N$.
Since this can happen for the univariate case, you should be able to see that it can also happen for the multivariate case where we have a marginalized posterior instead.
