If your different chains are sitting in different regions of the parameter space, averaging/concatenating them isn’t a good idea. The whole goal of MCMC is to get each chain to be sampling from the posterior. The “only” reason we run more than 1 chain is to actually test this assumption*. If 2 chains are in different parts of the space, then least least one of them (possibly both) is failing to achieve this. So this points to something more fundamental problem and I would expect the model needs to be substantially changed to address it.
*That’s a bit of an oversimplication.