As I mentioned before, the more appealing solution would be to drive the clocks to each of 64 T-flipflops to each of 64 relay drivers in order to turn the relays on and off with an 8x8 matrix. That would make them push-on/push-off types, merely requiring memory to record their state. Push-on/Push-off relays were once used in the telecom industry.

Another approach, given that the MCU had little else to do, would be to set the relay state from the matrix, retrieving the state from an in-memory buffer, and update it frequently enough that the state couldn't change due to the relay's behavior. That, too, would require thorough understanding of the relay characteristics, however. Such an approach would make the one-shot I previously mentioned desirable, though.

Given that I had to do this, I'd be tempted to use a CPLD with lots of pins to demultiplex the controls. That would even allow for a serial interface, which wouldn't consume so many pins. Additionally, using that sort of device would enable one to map the entire thing into external memory. That would allow for a really simple software interface, giving each relay its own memory location in the external memory map.

RE