This paper presents a frequency-shaping (FS) neural recording interface that can inherently reject electrode offset, 5-10 times increase input impedance, 4.5-bit extend system dynamic range (DR), and provide much more tolerance to motion artifacts and 50/60 Hz power noise interferences. It is supposed to be more suitable for long-term brain-machine-interface (BMI) experiments. To achieve the mentioned performance above, the proposed architecture adopts an auto-zero offset calibration to avoid system saturation, a delayed-signaling noise cancellation to attenuate kT/C noise, and an automatical data-splitting technique to reduce input-referred noise at low frequencies. Measured at a 40 kHz sampling clock and ± 0.6 V supply, the recorder consumes 50 μW/ch, including 22 μW for FS amplifier, 12 μW for gain-stage amplifier, 12 μW for buffer, and 4 μW for successive approximation register (SAR) analog-to-digital converter (ADC). The designed SAR ADC achieves an effective-number-of-bit (ENOB) of 11-bit in a 160 kHz bandwidth. In addition, the recorder has a 3 pF input capacitance and 15.5-bit (11-bit+4.5-bit) system DR due to the utilization of FS technique. The designed chip occupies 0.76 mm2/ch in a 0.13 μm CMOS process.