Sessa targets long-context memory retention

This stands out because the paper is not just claiming better long-context performance empirically, but arguing for a different memory mechanism with explicit theoretical advantages. The core differentiator is that attention is embedded inside recurrence, so past information can flow through many attention-mediated paths instead of a single attention read or a single recurrent chain; the main caveat is that the results are framed under explicit assumptions and matched regimes, so the real test will be whether the theory holds up under scaling, training stability, and broader benchmark coverage.