They replaced it with FastMM, which is a highly optimized iteration of RecyclerMM-like allocation strategy.

http://sourceforge.net/projects/fastmm/

The main issue with Borland's MM is that if in "best case" situations it can do well, in worst case situations it can become several orders of magnitude slower than other MMs, and can fragment to the point of running out of memory on a 2 GB machine with no more than a few dozen of MB allocated.