Memory latency is a correctness variable, not a performance one — 110ms round trips caused 7.2 of 12 redundant actions

arXiv 2607.05690·high signal

"Memory in the Loop" (arXiv:2607.05690, submitted July 6, revised July 19, 2026) moves memory read/write inside the agent's per-step loop rather than once per turn, which is only viable with an in-process store answering in ~100µs — three orders of magnitude below a network store. The headline number is a behavioral one: redundant actions were 0.0 of 12 at in-process speed but 7.2 of 12 at a 110ms cloud round trip, and recall went from 0/5 to 3.6–4.8/5. The practitioner move: pair an in-process vector store with a small *local* embedder (complete op ~40µs, store p50 80–165µs), since network embedding calls at 200–400ms become the dominant bottleneck and silently push your agent back into repeating itself.