Same may be necessary for values when we store them as nodes.

Possible solution:

After generating the value and reference nodes, put their hash in a "seen" array (as key).
Before generating the value and reference nodes, check whether they have been "seen" already.
The "seen" list would be a member of RdfBuilder, so no problems arise with changing options.

To reduce memory consumption of the approach I suggested above, use part (the first few digits) of the hash as the key in the "seen" array, keep the full hash as the value. For instance, using 4 digits would limit the size of the "seen" list to 2^16 entires.

When looking up x:

• !isset( $seen[ key($x) ] ) -> not seen
• isset( $seen[ key($x) ] ) && $seen[ key($x) ] === x -> seen
• isset( $seen[ key($x) ] ) && $seen[ key($x) ] !== x -> probably not seen

Finally, set $seen[ key($x) ] = x

Hat Tip to http://www.somethingsimilar.com/2012/05/21/the-opposite-of-a-bloom-filter/ and https://news.ycombinator.com/item?id=4251313

I'll try full hash first and see how much memory it consumes. If it's anything substantial then I'll go with what @daniel proposed.