AI agents embed text into vectors and compress them at scale. A compression is deemed acceptable when each item's nearest neighbors survive, yet agents also need to consolidate, delete, and compare memories—operations that depend on how far apart items lie. We evaluate both ends using two label‑free fidelity metrics from our prior work: local fidelity $L$ measured on the top‑10 neighbors and global fidelity $G$ measured on 30%–50% of the memory.
At 96 bytes per passage, rescored binary codes achieve $L=0.97$ but shift 10.1% of passages to a different theme; PCA with 8‑bit codes reaches $L=0.77$ and moves only 1.7%. Rescoring leaves the stored codes unchanged, so $G$ remains the same.
We prove that within the $G$ band, $1-\nu_k$ is bounded by at most $2$–$2.18$ times the square root of the near‑far rank‑violation rate, and that this rate follows the distance noise introduced by compression (correlation $0.979$). Leveraging this bound, we introduce locality‑aware quantization (LAQ), which turns the bound into a hinge objective on the codebooks of an OPQ index while preserving its byte size, rotation, and search pipeline. To our knowledge, LAQ is the first quantizer whose objective derives from a proven far‑end bound.
Reconstruction alone does not raise $G$ in any of 52 held‑out runs; the far‑hinge raises it in all 52. LAQ raises $G$ in all 52 runs ($p=8.5\times10^{-4}$ over 13 memory‑encoder groups), achieves the highest $G$ among seven equal‑size quantizers in all 12 Multi‑Doc‑2025 runs, scales to a million passages, and retains evidence recall.
Review: LAQ translates a theoretical bound into a practical optimization objective, delivering substantial global‑fidelity gains without increasing storage cost, and thus offers a robust quantization strategy for large‑scale memory retrieval.