<p>Recently, self-supervised learning (SSL) vision model has achieved great success, but its knowledge transfer usually needs an expensive parameter tuning cost. It hampers the broad application of SSL pre-training vision models. Some recent works attempt to extend the existing prefix prompt method to perform parameter-efficient tuning. However, these methods usually produce prompts and insert prompts in a model-agnostic way, and lack across-task information propagation pipelines. To address the above issues, we propose a Retrospective Prompt-Guided Parameter Tuning (<b>RPGT</b>) method to excavate the potential of contrastive self-supervised vision models on multiple downstream tasks. Firstly, RPGT dynamically generates retrospective prompt (RP) for contrastive SSL models in a model property-aware manner, and heuristically performs prompt insertion and interaction. It facilitates knowledge retrospection while avoiding the tedious prompt search process. Subsequently, a multi-task information propagation (MT-IP) module is proposed to construct across-task information propagation pipeline. It handles the knowledge sharing across tasks, accelerating the learning process of new tasks. Experimental results on twelve datasets demonstrate the effectiveness and generalization of RPGT. Our RPGT method can facilitate the transfer performance of various contrastive SSL models on multiple scenarios while significantly reducing parameter costs.</p>