Did you know? SWI-Prolog can parse SGML/XMLSearch Documentation:
SWI-Prolog owl logoHow to deal with the Prolog dynamic database?

The short answer is: if you can avoid using it, consider this first.

Why is the dynamic database evil?

The database is a non-logical extension to Prolog. The core philosophy of Prolog is that is searches for a set of variable bindings that makes a query true. It does this using depth-first search, binding variables to values. If an inconsistent variable binding is detected, the system backtracks: it finds the last choice it took, reverts all actions done since and continues with the next alternative. This is the core power of Prolog and gives it its logical basis.

Even if you cannot be bothered by logic, consider what this does for debugging: you can optimistically step over (skip) goals. If the answer is wrong, you hit retry, which takes the execution back to the start and you inspect the goal at a deeper level. In other words, the backtracking machinery provides you with a time-machine that allows you to go back in history.

Changes to the dynamic database are not reverted on backtracking, destroying all these nice goodies. Second, dynamic predicates are like destructible global variables and therefore subject to all the common problems with such objects: name-space pollution, state-full computation, initialization and cleanup requirements and lack of thread safety.

I need to preserve data over backtracking

Sometimes, results are computed and stored in the database to preserve them over backtracking. E.g.,



There are clean skeletons that deal with most of such cases. First, collecting all alternatives can be done using the all-solution predicates findall/3, bagof/3 or setof/3. For example:

young_persons(YoungPersons) :-
              (  person(Person),
                 age(Person, Age),
                 Age < 18

A second skeleton deals with aggregation of results obtained from backtracking. For example, finding the average age. Typically, one can use library(aggregate) for such problems:

average_age(Average) :-
                (   person(Person),
                    age(Person, Age)
                r(Count, Sum)),
      Count > 0,
      Average is Sum/Count.

I want to use dynamic predicates anyway!

Ok. You need to take care of some things.


If you want a dynamic database person(Name), your first problem is that it is initially empty and thus a call


yields the error below, which stops the computation

ERROR: Undefined procedure: person/1

A classical solution to this is to use the flag unknown, setting this to fail. Never use the unknown flag. It is a fossil that better remains buried, in particular for SWI-Prolog, where a large part of the development and library predicates are normally loaded on demand.

The correct solution is to use dynamic/1, as illustrated below. This directive tells the compiler that the predicate may have no clauses and also informs it that clauses may be added at runtime.

:- dynamic


If the dynamic database is used to store temporary results of a computation, you typically want them cleaned before you start. Here, never use abolish/1. Following the ISO standard, abolish/1 not only deletes all clauses, but also the (dynamic) predicate properties. Instead, use retractall/1:

cleanup_db :-

Cleanup may be done before the goal using it is started, but of course the disadvantage is that all useless results left at the end remain in the database. You can also do it at the end, but then you have to be careful that you actually reach the end. You may fail reaching the end due to unexpected failure or an error! One way to avoid that is using call_cleanup/2 as below. The call_cleanup/2 predicate guarantees that the 2nd argument is called, no matter how the first argument terminates.

go :-
      call_cleanup(go_with_clean_db, cleanup_db).

Take home

  1. Do not use dynamic predicates.
  2. If you need to preserve answers over backtracking, use findall/3, bagof/3, setof/3 or library(aggregate).
  3. If you really need dynamic predicates, use dynamic/1 to declare them, cleanup using retractall/1, and consider using call_cleanup/2 to ensure the database is cleaned.
See also
- multifile/1 if database predicates are loaded from a file.