Forth needs a big stack which means lots of SRAM.
Here is a wild attempt of writing a Forth Virtual Machine for the Arduino.https://github.com/mikaelpatel/Arduino-FVM120 primitives/instructions in the Virtual Machine. 3.5 Kbyte without the kernel dictionary strings (pure vm). Byte token thread with a built-in instruction level trace (optional). Multi-tasking with context switch back to the Arduino sketch on forth level yield, delay and halt. Cheers!