diff --git a/pkg/kubelet/cm/qos_container_manager_linux.go b/pkg/kubelet/cm/qos_container_manager_linux.go
index 4c29d23ce15..8f563b8c821 100644
--- a/pkg/kubelet/cm/qos_container_manager_linux.go
+++ b/pkg/kubelet/cm/qos_container_manager_linux.go
@@ -305,12 +305,6 @@ func (m *qosContainerManagerImpl) setMemoryQoS(logger klog.Logger, configs map[v
 		setUnified(v1.PodQOSBurstable, Cgroup2MemoryLow, 0)
 		kubeletmetrics.MemoryQoSNodeMemoryMinBytes.Set(0)
 		kubeletmetrics.MemoryQoSNodeMemoryLowBytes.Set(0)
-		// Clear per-pod memory protection only when MemoryQoS feature gate is
-		// enabled but policy is None (rollback scenario). When the gate is off,
-		// pods never had protection set, so there's nothing to reconcile.
-		if utilfeature.DefaultFeatureGate.Enabled(kubefeatures.MemoryQoS) {
-			m.reconcilePodMemoryProtection(logger)
-		}
 		return
 	}
 
@@ -330,37 +324,6 @@ func (m *qosContainerManagerImpl) setMemoryQoS(logger klog.Logger, configs map[v
 	setUnified(v1.PodQOSBurstable, Cgroup2MemoryLow, burstableRequests)
 }
 
-// reconcilePodMemoryProtection clears stale memory.min and memory.low on pod-level cgroups
-// when MemoryQoS is disabled or memoryReservationPolicy is not TieredReservation.
-func (m *qosContainerManagerImpl) reconcilePodMemoryProtection(logger klog.Logger) {
-	pods := m.activePods()
-	for _, pod := range pods {
-		podQOS := v1qos.GetPodQOS(pod)
-		var parentContainer CgroupName
-		switch podQOS {
-		case v1.PodQOSGuaranteed:
-			parentContainer = m.qosContainersInfo.Guaranteed
-		case v1.PodQOSBurstable:
-			parentContainer = m.qosContainersInfo.Burstable
-		case v1.PodQOSBestEffort:
-			parentContainer = m.qosContainersInfo.BestEffort
-		}
-		podCgroupName := NewCgroupName(parentContainer, GetPodCgroupNameSuffix(pod.UID))
-		podConfig := &CgroupConfig{
-			Name: podCgroupName,
-			ResourceParameters: &ResourceConfig{
-				Unified: map[string]string{
-					Cgroup2MemoryMin: "0",
-					Cgroup2MemoryLow: "0",
-				},
-			},
-		}
-		if err := m.cgroupManager.Update(logger, podConfig); err != nil {
-			logger.V(4).Info("Failed to reconcile pod memory protection", "pod", klog.KObj(pod), "err", err)
-		}
-	}
-}
-
 func (m *qosContainerManagerImpl) UpdateCgroups(logger logr.Logger) error {
 	m.Lock()
 	defer m.Unlock()
diff --git a/test/e2e_node/memory_qos_test.go b/test/e2e_node/memory_qos_test.go
index e7c0f335b3f..553e3e781db 100644
--- a/test/e2e_node/memory_qos_test.go
+++ b/test/e2e_node/memory_qos_test.go
@@ -613,16 +613,23 @@ var _ = SIGDescribe("MemoryQoS", framework.WithSerial(), func() {
 			restartKubelet(ctx, true)
 			waitForKubeletToStart(ctx, f)
 
-			// Pod cgroup memory.low is reconciled by the periodic setMemoryQoS loop
+			// QoS-class cgroup memory.low is cleared by the periodic setMemoryQoS loop
 			// (periodicQOSCgroupUpdateInterval = 1 minute) plus kubelet startup time.
+			var burstableCgroupPath string
+			if cgroupDriver == "systemd" {
+				burstableCgroupPath = filepath.Join(cgroupRoot, "kubepods.slice", "kubepods-burstable.slice")
+			} else {
+				burstableCgroupPath = filepath.Join(cgroupRoot, "kubepods", "burstable")
+			}
 			gomega.Eventually(ctx, func() int64 {
-				val, _ := memqosReadCgroupInt64(podCgroupPath, cgroupMemoryLow)
+				val, _ := memqosReadCgroupInt64(burstableCgroupPath, cgroupMemoryLow)
 				return val
 			}).WithTimeout(2*time.Minute).WithPolling(5*time.Second).Should(gomega.Equal(int64(0)),
-				"memory.low should reset to 0 when MemoryQoS is disabled")
+				"burstable QoS memory.low should reset to 0 when MemoryQoS is disabled")
 
-			// TODO(sohankunkerkar): Assert container-level memory.high resets to max once
-			// CRI runtimes support Unified map in UpdateContainerResources.
+			// NOTE: Pod-level and container-level memory.low values persist after rollback.
+			// Pod-level: clearing via systemd SetUnitProperties interferes with other cgroup settings.
+			// Container-level: requires CRI runtime support for Unified in UpdateContainerResources.
 		})
 	})