When you hear two versions of the same event, you can either update the memory of the first version with information from the second (updating) or retain both separately (differentiation). Prediction errors, the deviation between expected and actual outcomes, are presumably involved in mediating the competition between these processes. However, the mechanisms underlying this competition remain unclear. Here, we present a Hopfield network model that aims to account for the results of a behavioral study of memory updating. Participants heard conversational stimuli that were modified relative to previously encoded ones. Recognition memory was better for original than modified dialogues, and more frequently encoded originals were less influenced by modifications. Further, recognition of modified conversations showed a u-shaped dependence on the modification level, seen as a proxy of the prediction error. The present model encodes "original" patterns with varying strengths as well as modified versions of varying differences. The network stores the stimuli sequentially, weighting new patterns by their dissimilarity (prediction error) to previously stored patterns. During retrieval, partial versions of original or modified stimuli cue retrieval from the network. If the retrieved pattern is close to the cue, it is judged as old, otherwise as new. We find that very similar patterns interfere with each other, merging into a single attractor, while distinct patterns are stored separately and retrieved robustly. More frequently encoded originals form stronger attractors and are less affected by interference. Retrieval of stored modified patterns leads to a u-shaped similarity curve for increasing differences between studied patterns. The model's results align with the experimental findings, predicting: 1) similar new memories interfere with previous ones (updating), while dissimilar ones are stored separately (differentiation), 2) these outcomes arise from the network dynamics without a decision-making module, 3) reduced memory accuracy for modified stimuli is due to interference from similar representations, leading to low-confidence responses, and 4) the encoding weight declines if similar pattern are already stored in memory. These findings offer insights into how our brain deals with divergent information about the same event, showing that prediction error-driven competition between updating and differentiation can be achieved through network dynamics.