You can have large queue sizes (larger than RAM size) and siberite would still consume small amount of resident memory. You basically don't need a separate server with decent amount of memory for it. You can also can get benefit from two-phase reliable fetch - if your client gets disconnected without confirming a message, the message will be served to another client (very convenient if you use amazon spot instances for your workers).
Note that this also means that messages can be delivered more than once and/or that the clients need to remember the messages that they processed. In some setups that can be a showstopper.
Reliable fetch is a feature, not a protocol requirement.
You can use simple 'get work_queue' command to just get a message, or you can use 'get work_queue/open', 'get work_queue/close' - two phase fetch if you need a reliable fetch. You can also use 'get work_queue/close/open' command to acknowledge previous message and read a new one.
Ok so you can switch between at-most-once and at-least-once guarantees. While nice to have both options in a message queue, my point still stands.
Each of these have trade-offs and the way it is architectured here, in the at-least-once case you will have to either remember all the processed messages or be prepared to process a message multiple times, whatever that means in your specific use-case.