For one thing, there's not much of an incentive to port hundreds of thousands of lines of working legacy code to a different language just to make it more readable. Its age isn't a relevant issue either - sure, MUMPS debuted in 1966, but InterSystems Caché and GT.M are still being actively developed. Other databases are pretty old, too - the first version of Oracle was written in 1978.

If Caché had no strengths, I would agree with you, but as a non-relational database, it's pretty good. Sparse associative arrays are the default data structure, so it's very popular in, e.g., medical applications, where you would want to be able to store thousands of different things, but any given patient will only need a few of them.