Using direct numerical simulation of forced isotropic turbulence at ${\mathrm{\text{Re}}}_{\lambda }\approx 1300$ and $\approx 433$, together with two independent Markov-in-scale-by-construction null surrogates, we show that the Markov–Einstein coherence length of the turbulent energy cascade is . Conditioning the gap-scan test on the local flow state reveals that intermittent regions of the inertial range carry a coherence length $\Delta r\approx 3$–$4$, while the quiescent cascade at mid-inertial scales recovers $\Delta r\approx 1.0$–$1.4$, matching the canonical estimate to within one grid step. Near the dissipation range this pattern reverses: the core of the increment distribution carries more memory than the extreme tails, consistent with the spectral bottleneck. As a consequence, the unconditioned coherence length coincides with that of the intermittent component, whose memory is the last to decay as the scale separation grows and therefore sets the coherence length, giving $\Delta r\approx 3.2$–$3.6$ at the inertial-range centers, approximately three times the canonical reference $\Delta r\approx 1$. An independent Chapman–Kolmogorov test confirms the state-dependence, and a subsampling test shows the excess is present already at the $\sim$${10}^5$ samples-per-scale regime of the earlier experiments. The pattern is Reynolds-number-independent over ${\mathrm{\text{Re}}}_{\lambda }\approx 433$–$1300$. The Markov approximation underlying the cascade Fokker–Planck equation and fluctuation-theorem analyses is therefore substantially more restrictive for the intermittent component of the cascade than previously assumed.