The problem occurs when a configuration diverges and sets common workflow pipe elements in different parallel flows, a situation can occur where the convergent activity fails to evaluate correctly. When this happens an overlay cache copy of the workflow pipe remains in the convergence cache, which ultimately builds up to a memory shortage. Because convergence is a new feature some of the existing configuration is causing some unexpected results.
Our suggestion to overcome this issue is to provide a timeout for the convergence cache to do a controlled garbage collection.
Clear as mud then! I think he is trying to say there is a timing issue when memory is short but I really cant say for sure.