We present a memory architecture for long-context language models built on one idea borrowed from voxel game engines: partition a continuous space into cells, store only the occupied cells, and look them up without scanning. The AAMT routing stack already encodes such a partition (the 4-bit Meji and 8-bit Odu lattices); we formalize it as a strict coarse-to-fine voxel hierarchy over the TERA cube — a hexadeca-tree in which Meji is recoverable from Odu by high-bit extraction and a Morton key packs both levels — and extend it with a learned sub-voxel residual indexed by a navigable small-world graph (HNSW). The result is a two-tier rolling memory whose effective context grows from a fixed working set to a million-item global store at constant attention cost. Measured on an Apple M1 Max (32 GB) against the platform's own corpora: a 64% relative gain in retrieval recall@10 from composable levers (bge-large embedding, PCA-whitening, Hopfield consolidation), full recall retention at a 285× scan-cost reduction on a 1M-item HNSW store, and an honest negative result on off-the-shelf cross-encoder rerankers. WP-20 extends WP-15 (Vortex-Addressed Semantic Memory) with the residual and graph index it left open.
Weslyn Cory Whitehead (Fri,) studied this question.