Noises exist in digital circuits can be leveraged as the source of entropy to design true random numbers generators (TRNGs), which is an important primitive component in cryptography and hardware-based security applications. In this paper, we present practical experiments and results of a TRNG implemented on magnetic random-access memory (MRAM) dies, fabricated by the thermally-assisted-switching MRAM (TAS-MRAM) technology. We, first, explain how one can find out a heating voltage value using that writing operations in the TAS-MRAM dies have a stochastic behavior. Then, we propose an improvement based on a feedback loop from being generated random bits. It helps to adjust the founded heating voltage value for writing operation with the aim of reaching the maximum entropy. Finally, we report the results of some post-processing methods, which are usually required in TRNGs to successfully pass the statistical test of NIST SP-800. The results show that one can generate random numbers with a high quality of randomness using the proposed TRNGs besides simple post-processing methods.