From the Forth-83 standard:

I assume that the compilation vocabulary is the vocabulary returned by CURRENT @.
My question is, if <name> is not found in the current vocabulary, can FORGET search the parent vocabulary(ies) of the current vocabulary for <name> or is it limited to only searching the current vocabulary?

Perhaps I should elaborate. My Forth's FORGET starts like this:

VOCABULARY is written such that vocabularies are not chained to their parent vocabulary but they do have the address of their parent in their second cell. This is to keep FORGET from finding <name> if it is not defined in the current vocabulary. The high level word FIND takes care of following the chains, but it might be a bit slow:

If the standard permits FORGET finding <name> in the current vocabulary or one of its parent vocabularies, I would like to rewrite VOCABULARY so that vocabularies are more like the the fig Forth model. I could then change FIND to:

By the way, 0EXIT is a code word that exits if the top of the stack is false
0EXIT ( FLAG -- )

This is why I would appreciate some clarification on the Forth-83 standard's meaning concerning FORGET.

[EDIT:] Problem solved with a little creative rewriting of the find primitive and adding a new primitive VFIND which only finds a name in a given vocabulary. VFIND has no parameter field. Its code field points one byte into (FIND)'s parameter field.

When FORGET is used, all words, including vocabularies, that are defined after the forget point ( word to be forgotten ) are forgotten as well. The vocabulary structure in my Forth is a tree. Forth is the root vocabulary. A new vocabulary's parent vocabulary ( the one searched after the new vocabulary is searched ) is the vocabulary where it was defined.
These are the features of FORGET in my Forth:
Sets CONTEXT equal to CURRENT incase the context vocabulary is forgotten. This could be changed so that only if the context vocabulary will be forgotten will CONTEXT be set equal to CURRENT.
Only searches the CURRENT vocabulary.
Prunes VOC-LINK so that the latest defined vocabulary that will not be forgotten will be the latest vocabulary in the link, the one VOC-LINK points to.
Follows VOC-LINK and prunes all remaining vocabularies.
If any of the kernel defered words has a 'vector' that will be forgotten, FORGET resets that defered word to its default value. This is handled with a table and a DO LOOP to keep it small.
It has an internal fence ( a literal in the definition of FORGET ) as well as the variable FENCE. The forget point must be above both of them or FORGET aborts with the message "PROTECTED".

I think this is a safe FORGET, but can it be made better?

A little information about EMPTY, the word to reset the dictionary to its bootup state, or smaller.

There is a place in the definition of FORGET just after the word to be forgotten has been found, the word's link field address is on the stack. This Forth's EMPTY branches to that address after doing the following. It fetches the 'empty address' from a location in the boot area and takes the unsigned minimum of it and HERE. It then stores a copy of that value in the 'empty address' location. Finally, it executes FORTH DEFINITIONS before branching to FORGET with that value on the stack.

Any suggestions? Ideas for improving both FORGET and EMPTY?