I think a context (an active unconscious process influencing current actions and perceptions, the stuff you are doing at the moment, in whatever way) may best be implemented as a state machine. This way, it is easy to see how a process may be interrupted by another process, and later be continued.

The knowledge about how to do things is stored in state machines. These state machines are inactive, they are not instantiated. They have initial states and final states. An initial state is not really a state (it is a pseudo state, in UML lingo). An initial state is connected to the first real state to the FSM by a transition, caused by an event. If the event occurs, the FSM goes from the initial state to the first state. But what's more important: the FSM is instantiated at that time. That is why the initial state is special. The same holds for the final state, but reversed: the FSM is removed.

In this implementation there needs to be a daemon (a permanent background process) active that takes care of the recognition process: check if the incoming event is one that causes an "initial transition", and if so, instantiate the FSM. Once the FSM is active, events will be sent to the FSM directly. For consistency, the daemon must also remove the instances if a "final transition" occurs.