What is the best way to track down a memory leak that is only found on one customer\'s test/release box, and no where else?
It's either code, data or configuration.
Since you say the code is not faulty 100% of the time, I would blame configuration. Take a copy of the configuration (and optionally some data) and try to replicate the problem; you won't know you've found and fixed it without reproduction.
Finally, solve it with a memory profiler.