Hello Wanlu,
I think this a normal behaviour. Temperature is thermostatistical property, then its significance is "real" when you use an average over time or particles. When you run a MD with hundreds of particles, you have a particle average at each frame, but when you use a single particle, the temperature is not well defined for each frame, but the time average temperature its better defined. As an example, a single particle harmonic oscillator, you define an initial and conserved total energy, the particle changes their potential and kinetic energy along the parabola, at the left and the right "all" (forgot the zero point energy) the energy is potential energy, and in the middle "all" is kinetic, the kinetic energy (and the particle temperature) oscillate from total energy (the highest temperature) to zero in an harmonic framework, and the time average gives you the thermostatistic temperature. If you couple a thermostat to regulate the temperature, you will add and subtract kinetic energy depending the time local situation, and never achieve the equilibrium.
This is what happens in a single molecule MD (with some differences from the previous picture, due to the oscillators are not totally harmonic, and the time integration is not infinitesimal).
If you want to simulate a single molecule at a particular temperature T*, I recommend to run a NVE MD, with an initial temperature of 2T*, and a minimum energy geometry (at the same level of theory). In this framework, from the equipartition principle, half of kinetic energy goes to the potential energy (the average values) and the average temperature would T*, or very near (due to the system and the simulation are not ideal or perfect). This is like as running a harmonic MD with an initial geometry in the bottom of the parabola, and giving some kinetic energy (where kinetic energy is equal to total energy at the beginning).
Regards!