Only for cross-agent memory?

[bitnom]

The readme says:

Skip it if you…

only use a single provider's native compaction and don't need cross-agent memory

Why would this be true? If it's doing code/text compression and rtk, why would I need to be sharing memory between multiple agents to see the token savings? Is it running against every request? Given this "headroom learn" feature, the quote is making even less sense.

thanks in advance for clarification