The knowledge network can transmit singular values, ranges of alternative values (including numbers, objects and strings), lists, but sometimes these are not enough. If the object is complex, such as a surface, it can't easily be parcelled up and sent. An alternative method is to transmit activation. An operator with control changes the environment for another operator, then transfers control to that operator. The operator comes up, examines its environment, it may alter its internal state, and then either relinquishes control (it does nothing) or passes control to another operator. In this way, an arbitrarily complex message may be transmitted through the network, although at no time did any point in the network transmit all the details of the message.
As an example, consider distributions on variables and relations among variables based on those distributions. We have a single distribution on a variable and many relations with other variables. The distribution has to change in response to changes in distributions on the other side of those relations. We cannot transmit a message through the relation because the distribution is the result of all the relations, not just one. Instead of transmitting a value, we give control to the distribution operator by placing it on the activation queue. The operator's job is to access all changed relations and change the distribution it holds internally. If the range of the variable changes, this is sufficient message to transmit to the relations around the variable, and to any other operators that are interested in the distribution, such as a Bayesian Less Than operator. If the range hasn't changed, we need to give control to the operators which can directly access the distribution.
Another example concerns booking of resources. A number of usage operators have tentatively booked a resource, where the bookings overlap and exceed the availability of the resource. If a hard booking is then made by one usage operator, it has to "bump" the other usage operators which now cannot overlap it. The booking is made, the other operators gain control and change their bookings, potentially bumping other usage operators in the particular resource and in other resources as they do so. There was no single message that could be sent, the assemblage being too difficult to describe except by allowing it to behave as it is constrained to do.
How is this different to seeing a change on a variable in a notional network, then giving control to each of the constraints on the variable? Such an arrangement implies that there is a data structure at the variable capable of storing whatever message turns up. The operator is combining many messages at once to build something more complex than any of them, but never saves the complex message into any single data structure. It may save part of the message it develops internally, but it also passes control to other operators which may be interested.
"The activity is the message." No one point in the network can be isolated and the message extracted. This is surely what we would expect in a complex system. The network provides the framework for the transmission of activity and in that way guides what activity takes place. The activity can take the form of a cascade, as one usage operator spreads its effects into others through a resource, or a spreading activation, as one distribution re-establishes equilibrium with those around it through relations.