This looks pretty interesting to me. But it leads to a lot of questions that are unanswered on the webpage.
The claims of 30% faster for a particular workload are reasonable, but what configuration is needed and what exact workload are we talking about? I'd love to see some benchmarking code.
What are the limitations imposed by running in memcached mode? Can you still do clustering?
What other changes are they planning? Site just says "many". Are we talking about memory allocation or persistence changes? Maybe additional interface compatibility modes? It's hard to bet on this when we don't have any idea what else it will bring to the table.
Where possible, I'd really like to see these improvements factored into distinct redis modules and made available as plugins to the core.