A generalized version of a memory automatic repeat request (ARQ) scheme [15-221 using hard, soft, and completely soft detection is examined. The communication channel is modeled as an additive white Gaussian noise (AWGN) channel. The following six combining schemes are considered. In Scheme 1, error detection is performed on each of the n received copies. In Scheme 2, all n copies are combined and error detection is performed on the resulting packet. Scheme 3 uses Scheme 1 followed by Scheme 2 (if decoding using Scheme 1 is not successful). In Scheme 4, each incoming packet is combined with the copies received so far, and the resulting combined packet is checked for errors. Scheme 5 uses Schemes 1 followed by Scheme 4. The last scheme, Scheme 6, attempts decoding on up to all 2n- I possible combinations of the received copies. For an arbitrary detector quantization, analytic expressions for the retransmission probability, PϜ of Scheme 3 are derived, and numerical procedures for evaluating the PϜ of Schemes 4, 5 and 6 are described. Numerical examples showing the dependence of PϜ on various parameters, such as the signal to noise ratio (SNR), number of transmissions n, number of quantization levels M and packet length L are given. Although Scheme 6 has the lowest PϜ its decoding complexity increases exponentially with n. The six combining schemes are applied to the Weldon ARQ scheme [13,14,22] It is found that the throughput T is generally improved using multiple copy combining schemes. The throughput can be further increased by using a soft detector. For block error rates BKER > 0.5 and soft detectors, the throughputs of Schemes 2 to 6 are typically higher than that of an ideal selective repeat ARQ system with no packet combining. In general, Scheme 3 appears to be a good choice among all the six schemes with a good performance and a simple decoder structure.