The brain of the fruit fly is well suited to study general principles of brain function, because it is small enough to be studied at the single-cell level and its connectivity is well known. This makes it possible to build mechanistic models with the same connectivity as the fruit fly brain, and to train them on tasks in order to constrain the free model parameters. It has recently been shown that such connectome-constrained ‘deep mechanistic network’ models (DMNs), when task-optimized to perform motion detection, can predict the tuning of single-cells in the fruit fly visual system and can be used to generate hypotheses about the mechanisms of neural circuit function [1]. For many neurons in the fruit fly visual circuit, single-cell activity has been experimentally recorded. We here ask whether it is possible to integrate such neural activity measurements into DMNs in order to further constrain them and enhance their predictions of computational mechanisms. At the same time, our goal was to maintain the network’s ability to perform the task it was originally trained on, in our case motion detection. We trained ensembles of connectome-constrained models using both sequential and joint training approaches to match simulated activity measurements and detect motion in naturalistic video sequences. In sequential training, models were first trained to match neural activity measurements and then to perform motion detection, or vice versa. However, we found that this approach resulted in models forgetting the first skill when subsequently trained on the second. In contrast, joint training, where the models were simultaneously trained to match neural activity and to detect motion, resulted in models that performed well in both tasks without forgetting either. We observed that models trained to match measurements from only a few cells also better predict the activity of other cells for which no measurements were available. Our findings suggest that incorporating neural activity measurements refines connectome-constrained and task-optimized DMNs, enhancing their consistency with experimental data and improving their predictive capabilities.