Currently， the majority of structure damage identification methods based on deep learning rely on deep neural networks to automatically extract damage-sensitive features of structures and achieve pattern classification recognition through the differences in features between damage states. However， these methods face challenges in the accurate quantification of damage and require a large amount of data for model training. This article proposes a damage detection method based on a model-embedding recurrent neural network （MERNN）. Firstly， a data-driven convolutional neural network was used to establish the mapping relationship between load and response. Then， the traditional recurrent neural network was improved using the Runge-Kutta method to create a numerical computing unit based on the recurrent neural network architecture. Finally， based on the loss function composed of the residual errors between measured responses and computed responses， the structural stiffness parameters were updated with the automatic differentiation mechanism of the neural network to achieve structural damage identification. Damage identification results of a numerical three-layer frame and a laboratory-scale shear-type frame indicate that the proposed method can accurately quantify structural damage based on the limited amount of response datas.