To quibble about one more thing, multitasking doesn't increase memory requirements significantly on anything big enough to run an interpreter. On this amd64 machine under Linux, sizeof(jmp_buf) is 200 bytes; that's how much space each new task would needs, plus of course the size of its stack. If I build for i386, it's 156 bytes. But of course you can define the calling convention to reduce this: on 16-bit Forth systems a suspended (cooperative!) thread was typically a data stack pointer, a return stack pointer: 4 bytes. (But typically each thread had its own dictionary of at least 4K.) Preemptive multitasking generally requires at least the size of the register file, since you can't discard caller-save registers on every context switch.

What does increase memory requirements enormously is memory protection — not mere multitasking but allocating exclusive segments or pages to each new program. That's a big part of the appeal of systems like Oberon, Smalltalk-80, and of course early Oak/Java: by implementing memory protection at the language level rather than the hardware level, you could get very fine-grained memory sharing without endangering system stability.