Threshold-type resistive switching (RS) is essential for selector devices and electronic neurons [1], but Ag-based diffusive memristors are typically demonstrated on large-area devices (>1 µm2) fabricated on unfunctional substrates (SiO2/Si), limiting their relevance for integrated systems [2]. In particular, switching behaviour, endurance, and operating conditions do not scale predictably with device size or integration environment [3-4]. Here, we demonstrate Ag-based threshold switching devices using multilayer hexagonal boron nitride (h-BN) integrated at the back-end-of-line (BEOL) of CMOS microchips, with device areas down to ~0.05 µm2. While large-area Ag/h-BN/Ag devices fabricated on SiO2/Si exhibit forming-free volatile switching at low VSET (~0.47 V) and high endurance up to ~9×106 cycles (Fig. 1), direct on-chip integration leads to the loss of threshold switching due to increased energy required for filament formation in reduced device dimensions (Fig. 2-3). Removing the Ag layer, results in the absence of RS completely, and only erratic resistance fluctuation is present (Fig. 4). We show that this limitation can be overcome by engineering the top electrode composition. By introducing a limited-volume Ag layer (~1 nm) within the top electrode, we control the formation of conductive nanofilaments and recover stable volatile switching in fully integrated 1T1M cells. The optimized devices exhibit forming-free operation, switching voltages below 1.5 V, and endurance up to ~4 × 106 cycles (Fig. 5). These results demonstrate that switching behaviour in diffusive memristors is strongly dependent on device scaling and integration conditions, and cannot be extrapolated from large-area devices. Finally, we validate the use of these devices as leaky integrate-and-fire neurons in spiking neural networks, achieving correct classification in 94% of cases. This work establishes a pathway for integrating 2D-material-based threshold switching devices into CMOS platforms for neuromorphic computing.