The memory capabilities of large language models (LLMs) have attracted increasing attention recently. Existing retrieval‑based memory methods usually treat all memories as homogeneous and apply a single retrieval strategy, which limits performance in heterogeneous scenarios. To address this, the paper introduces a multi‑class memory dataset called TriMEM, providing precise annotations for various memory types across diverse contexts. Building on TriMEM, the authors propose the MemoType framework: a learned router model automatically identifies the type of each memory and the corresponding query, then selects a tailored retrieval strategy. By routing both memories and queries, the system can employ the most suitable retrieval method for each memory class, improving overall precision. Theoretically, it is shown that any single retrieval strategy has an upper bound on its expected precision in a multi‑class corpus, causing systematic degradation. Experiments on three public datasets demonstrate that MemoType achieves up to a 16.18% gain in Recall@1, consistently outperforming prior approaches.
Review