As the proverb goes “two heads are better than one” and “everybody’s business is nobody’s business”, the phenomenon of crowd intelligence can be easily observed in our daily lives. At the same time, with the rapid development of network technologies, crowd intelligence becomes much more complicate and universal, for human, enterprises, governments, equipments and articles turns to be more and more intelligent, and these intelligent agents are connecting to form numerous crowd network systems, such as e-commerce platforms, networked supply chains, Wikipedia and network elections (Michelucci and Dickinson, 2016). As main mode of modern service industry and future economy society, the research on crowd network can greatly facilitate governances of economy society and make it more efficient, humane, sustainable and at the same time avoid disorders (Chai et al., 2017). However, because most results cannot be observed in real world, the research of crowd network cannot follow a traditional way. Simulation is the main means to put forward related research studies. Compared with traditional large-scale simulations, crowd network simulations have several obvious challenges as follows: � Dynamic. Member attributes and states of crowd network simulations may vary with time in an uncertain mode. Members are more loosely coupled, but member behaviors and intention variations may lead to change of group states and intentions in extending scopes. � Diversification. Time advance strategy may base on slow variables, events, clock or hybrid mode. And as members are multiform and multi-disciplinary, transactions are uncertain and various, disturbances have several sources and subscriptions exist in different layer and aspects, disturbances injection strategy and matching strategy are all need to take diversification into considerations. � Scale. Member of crowd network simulations may need to achieve millions or even trillions so as to discover or verify its principals and regularities. In a word, crowd network simulation is a new development of large scale simulations, which faces both opportunities and challenges. This paper proposes a novel reflective memorybased architecture to resolve problems mentioned above, and the rest of this paper is organized as follows: Section 2 gives a brief review of related efforts towards crowd network simulations, Section 3 analyzes shortcomings and advantages of high level architecture (HLA) based simulations, Section 4 proposes a novel reflective memory-based simulation architecture; Section 5 demonstrates an implementation architecture and Section 6 draws a conclusion of the proposed approach.