Emotion Recognition for Conversations (ERC), a key research area in Natural Language Processing (NLP), has been increasingly implemented in human-computer interaction (HCI) systems. Existing approaches employ graph-based or sequence models to extract features from utterances, yet these features are often insufficient for emotional reasoning tasks. With the growing adoption of Large Language Models (LLMs), researchers have integrated speaker identification techniques to boost model performance. However, the hidden information about the current dialogue is not fully exploited. In this work, we introduce a novel framework enabling large language models to deeply mine context and speaker hidden information (DM-ERC). First, speaker latent commonsense is generated by Llama2-Chat from historical utterances and injected into the large language model to improve its emotional dynamics understanding. Next, emotion prediction experiments are conducted on the pre-trained LLM to anticipate future emotional trends. Furthermore, a similar utterance retrieval module is integrated to assist reasoning during emotion recognition. Extensive experiments validate that the framework outperforms current baselines on three publicly available datasets.