From what I'm understanding, this format uses a variable number of bits, automatically growing and shrinking as needed. How can this be implemented efficiently

From what I'm understanding, this format uses a variable number of bits, automatically growing and shrinking as needed. How can this be implemented efficiently? If there's a cap on the maximum number of bits (eg, 64), why not just use that maximum number all the time, and not have to worry about shuffling things around (bitwise!) in memory?