Memory Grows. Context Does Not Have To. (ExergyNet: Persistent Machine Infrastructure for the 100M-Token Horizon)

The practical challenge in scaling autonomous machine intelligence is not model capability—it is infrastructure. As agents accumulate persistent histories, prior approaches have treated the growth of stored context and the growth of active model-facing inference context as the same problem. They are not. This paper measures the distinction. Across a sealed development campaign scaling from 32,000 to 101,152,694 estimated tokens, ExergyNet’s externalized persistent evidence architecture (xLMP) held active model-facing context (K) bounded at a median of 800 tokens, while the underlying corpus grew by more than 3,100x. The state-to-active-context ratio at the 100M-token operating point was approximately 125,000:1. There was no statistically significant positive trend between corpus growth and context growth within the tested envelope. Furthermore, we demonstrate that legacy full-corpus retrieval scanning presents a severe computational bottleneck at scale, masking GPU efficiency behind CPU idle time. By implementing an O(vocab) morphological prefix-expansion index (LNES-84), retrieval latency at the 100M-token scale collapsed, yielding a 220.5x median and 415.9x P95 paired retrieval speedup over the baseline, with zero right-to-wrong answer regressions across 190 matched queries. Finally, we formalize LNES-22, an out-of-band cryptographic authority control plane that mathematically decouples a model's intelligence from its execution authority. This architecture was physically validated via FAA Exemption No. 26214 for a 275-lb autonomous heavy-lift aircraft. We conclude that persistent state, semantic routing, and execution authority must be established as infrastructure independent of the reasoning model.