In this paper, we propose a deep state-action-reward-state-action (SARSA) $λ$ learning approach for optimising the uplink resource allocation in non-orthogonal multiple access (NOMA) aided ultra-reliable low-latency communication (URLLC). To reduce the mean decoding error probability in time-varying network environments, this work designs a reliable learning algorithm for providing a long-term resource allocation, where the reward feedback is based on the instantaneous network performance. With the aid of the proposed algorithm, this paper addresses three main challenges of the reliable resource sharing in NOMA-URLLC networks: 1) user clustering; 2) Instantaneous feedback system; and 3) Optimal resource allocation. All of these designs interact with the considered communication environment. Lastly, we compare the performance of the proposed algorithm with conventional Q-learning and SARSA Q-learning algorithms. The simulation outcomes show that: 1) Compared with the traditional Q learning algorithms, the proposed solution is able to converges within \myb{200} episodes for providing as low as $10^{-2}$ long-term mean error; 2) NOMA assisted URLLC outperforms traditional OMA systems in terms of decoding error probabilities; and 3) The proposed feedback system is efficient for the long-term learning process.