A quick-hack with b_setval/nb_setval to achieve backtrackable asserts, because I coulnd't find one. If you're familiar with how the assert/retract behaves, this library should work how you expect it to ( with backtracking resetting the assert/retracts, ofcourse). But I still recommend you read the whole readme for the caveats.
The library can be used in programs using :- use_module(backtrackable_asserts)..
The library's export list is the following: [b_assert/1, b_asserta/1, b_assertz/1, b_retract/1, b_query/1].
They're named so copying the b_ in backtrable versions of predicates (e.g. b_setval).
b_assert* & retract work similar to the corresponding assert* & retract predicates. The notable difference is b_query - All predicates which are asserted using b_assert cannot be queried directly. They must be queried using b_query. **Do NOT try to mix regular dynamic facts and b_assert dynamic facts. It will not work. **
?- use_module(backtrackable_asserts). % Include the module
true.
?- b_assert(foo(a)), foo(X). % these facts cannot be queried directly.
ERROR: Undefined procedure: foo/1 (DWIM could not correct goal)
?- b_assert(foo(a)), b_query(foo(X)). % You must use b_query/1
X = a.
?- assert(bar(b)), b_query(bar(X)). % Conversely, you can't query regular dynamic facts with b_query either.
false.
?- b_assert(foo(a)), ((b_assert(foo(b)), fail); b_query(foo(X))). % The assert to foo(b) was removed on backtracking.
X = a.
?- b_assert(foo(a)), b_assert(foo(b)), (b_retract(foo(a));true), b_query(foo(X)). % this time with retract
X = b ;
X = a ;
X = b.
If you asserting and retracting the same fact over and over is not going to be efficient. For a query (or retract), We'd have to 'walk down the stack' till we find a fact that hasn't been retracted yet. This would be of the order of the number of facts matching your query that have been asserted. Otherwise, I expect the overhead to stay reasonably small.
Because I can imagine it being useful, nb_assert provides a way to persist facts which aren't reset on backtracking - The difference to regular assert being that it is now queriable using b_query. So you can mix nb_assert and b_assert.
How do they mix? b_retract on an nb_asserted fact does retract it. The retraction will reset on backtracking, the assert will not. To get rid of it completely, do an nb_retract.
Use nb_retract carefully. It does a full retract of an (n)b_asserted fact.
nb_retract'ing an already retracted fact will fail. The reasoning is: Which foo does nb_retract retract?
nb_assert(foo), b_asserta(foo), b_retract(foo), nb_retract(foo), b_query(foo).
Hence, nb_retract always retracts the top-most occurence. The following query succeeds once when the b_retract is backtrack-resetted.
nb_assert(foo), b_asserta(foo), nb_retract(foo), (b_retract(foo);true), b_query(foo).