continuity-layer

Infra is a provider-neutral continuity layer for AI agents that preserves authoritative state across interruptions, reducing model-token use by 49.4% and accelerating recovery by 60.8% in internal controlled tests.

Position paper proposes a “continuity layer” that preserves what models learn over time, introducing Decomposed Trace Convergence Memory and the ATANT benchmark to measure 100% isolated, 96% cumulative recall on a 250-story corpus without language models in the loop.