So the number '123' is stored in sequential memory locations: x, x+1, x+2 with a word mark under the '1'.
To add (assembly: A X,Y) e.g. '456' to said number at y, y+1, y+2 (in this case the 4 need NOT have a word mark (but usually does) because a word mark in either field ends the operation) the machine starts at the high-order address -- so it starts at address x+2 and y+2, develops any carry, applies that to x+1 and y+1, adds, and finally moves to the hundreds digit, adding any carry from the tens. (The destination address holds the sum in this case, '456' goes to '579')
Also, to move '123' somewhere, you specify the address of the '3' and the field is moved right-to-left to the destination (decrementing the internal pointer, and stopping when finding a word mark).
So, it more or less was Big Endian, but was still able to use a serial ALU by starting at the highest address!
The only exception is IBM, which used both BE bit and byte numbering. (Bit 0 is the most significant bit, and its value depends on the width of the data in question --- exposing the major disadvantage of BE, i.e. a length-dependent term.)
An old adage I've heard and found useful: "Little endian is Logical; Big endian is Backwards."