Forth is an imperative language and as such you will have to model memory state (at least) somehow, if you want to use purely functional language as a representation. But that's the cost of doing business.

The thing is though, you don't translate to LC for performance, but for understanding. At any point, the LC evaluator might recognize a known subterm and replace it with a known equivalent. Depending on the goal, that might help to improve evaluation performance (skip known steps of beta reduction), reduce memory consumption (reduce size of the term), or improve readability (refactor code to use better known abstractions).