Defect-Engineered Floating-Gate Synapses With Programmable Relaxation Dynamics for Multi-Timescale Neuromorphic Computing.

The human brain processes temporal information across timescales spanning milliseconds to years through hierarchical memory systems, the biological capability that remains largely inaccessible to artificial neuromorphic hardware. Here, we present synapses with programmably tunable relaxation dynamics to emulate such multi-timescale cognition. By precisely engineering defect density, we achieve systematic modulation of charge trapping/de-trapping kinetics, enabling a continuous transition from nonvolatile to volatile memory within the synaptic device. The transistors exhibit a full spectrum of neuromorphic functionalities, including paired-pulse facilitation, multilevel conductance states, and relaxation times tunable across orders of magnitude. Notably, we introduce a parallel dynamic memory superposition architecture, which integrates devices with complementary timescales into a physical reservoir computing framework, enabling simultaneous extraction of both long-term periodic patterns and short-term transient fluctuations without signal entanglement. The heterogeneous reservoir demonstrates superior performance in computational tasks such as multifrequency oscillator prediction, significantly surpassing single-timescale counterparts. Our work establishes defect engineering as a compelling paradigm for programming temporal dynamics in neuromorphic systems based on two-dimensional materials, offering a viable pathway toward energy-efficient and adaptive edge intelligence for real world time-series processing.