StockMem is a novel event-reflection memory framework designed for stock price prediction using large language models (LLMs). It addresses the limitations of general-purpose memory architectures in financial domains by converting unstructured online news into structured, temporally-aware event knowledge. The framework consists of two main components: an event memory that extracts, merges, and tracks events over time to capture incremental information (ΔInfo) reflecting market expectation discrepancies, and a reflection memory that learns causal relationships between event sequences and stock price movements, forming a retrievable experience base. During inference, the system retrieves historically similar event sequences using a hierarchical similarity metric (combining type-level and group-level Jaccard coefficients) and LLM-based filtering, then synthesizes recent events, incremental information, and retrieved experiences to predict price direction.

Experiments on four technology stocks from the CSMAR database demonstrate that StockMem consistently outperforms state-of-the-art memory-augmented models (FinMem, MemGPT, MemP, Zep) in accuracy and Matthews correlation coefficient. Ablation studies confirm the importance of structured event representation over text summaries or cluster opinions, the critical role of incremental information from event tracking, and the effectiveness of cross-company historical pattern matching. The framework also provides strong interpretability by clearly tracing the information chain affecting stock prices, enhancing transparency and trustworthiness in financial decision-making.