In MP3:

CBR (and also to some degree ABR) uses a constant defined bitrate. Because that bitrate is taken into consideration at every frame, there will be certain moments of such complexity that they can't be properly encoded within the limitations of the chosen bitrate; they need a higher bitrate than the defined one. Therefore, the MP3 spec defines a bit reservoir.

Example: a certain moment in a song needs 130 kbit to be accurately encoded (as defined by the psychoacoustic model and the settings of the encoder) the CBR bitrate is set to 192 kbps. 62 bits are not used (192 - 130 = 62). those bits can be used for the next frame. This also works with consecutive frames, until the point where the bitreservoir has the size of a 320 kbit frame.

With VBR, the encoder can choose the needed framesize at any moment, again as defined by the psymodel and the quality settings. That's why VBR (e.g. in LAME) doesn't use bit reservoir.