The Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) will observe active galactic nucleus (AGN) sky densities of ∼ 1000-4000 mathrm deg ^ -2 , enabling massive photometric reverberation mapping from the continuum light curves. We develop a meta-learning framework for photometric reverberation mapping of AGNs in large time-domain surveys. We present the framework based on attentive latent neural processes (ALNPs), designed by the SER-SAG-S1 directable software in-kind team to the LSST, to perform data-driven quasar photometric reverberation mapping for massive time-domain surveys. The framework consists of clustering AGN light curves with similar topologies in each photometric band using self-organizing maps, followed by a novel combination of ALNP and mixture density models for unsupervised learning of light-curve structures, the underlying physical parameters of the supermassive black holes (SMBH) that power the AGN, and information on the transfer functions of their accretion disks. We conducted experiments on simulated AGN light curves with varying cadences and transfer functions, and real data from the Transient Facility (ZTF). The key results are that (i) the latent space of the ALNP encodes information on the transfer function and SMBH parameters, (ii) the light curves are reconstructed with improvements of 60-70% over other regressors (e.g., Gaussian process models) trained over an ensemble of light curves, (iii) the transfer functions are reconstructed with an improvement of ∼ 35% over the training prior in our chosen low-variability cluster, iv) intrinsic SMBH and light curve red-noise parameters are recovered with an improvement of ∼ 34% over the training prior, with an improved reconstruction at lower values for most parameters and higher values for the redshift, and v) the framework can be applied to ALNP representations of real light curves after it is trained on simulated datasets. Zwicky The capability of ALNP to capture diverse transfer functions and SMBH parameters enables us to integrate various trained ALNPs into an ensemble or to make improvements through latent space clustering with SOMs. This allows the framework to handle versatility and potential novelty in transfer functions, making it well suited for diverse and unseen AGN data from upcoming large-scale surveys.