The focus of this paper is on the management of distributed knowledge, in environments where a large number of autonomous and heterogeneous systems collaborate to support cer- tain tasks. All information is represented as logical facts, and tasks correspond to logical rules. We use the language WebdamLog, a recently introduced variant of datalog for dis- tributed data management. The main novelty of WebdamLog compared to datalog is its use of delegation, that is, the ability for a peer to communicate a program to another peer. This paper introduces a WebdamLog inference engine. Our main contribution is to demon- strate that such an engine may be implemented with good performance for very dynamic applications, i.e., for applications with rapidly evolving knowledge and with heavy knowledge exchange. To avoid reimplementing standard features of datalog engines, we developed Web- damLog on top of the Bud system, which has already been demonstrated to be very efficient. In this paper, we describe our implementation of the features that are specific to WebdamLog, and include: (1) support for complex rules with predicates in the rule body that are evaluated by remote peers; (2) support for variables for peer and predicate names that allow for very flexible management of distribution, e.g., for the discovery of new peers; and (3) management of delegations that lead to changes in the programs of remote peers. Our focus on dynamic peer state and on rapidly evolving knowledge leads to new perfor- mance challenges, prompting us to introduce new evaluation techniques. A key component of our approach is a novel kind of a provenance graph that allows for performance optimiza- tions in this dynamic context. We describe the provenance graph and its use for performance optimizations, and outline experiments that we are conducting.