Reproduce “AssoMem: Scalable Memory QA with Multi-Signal Associative Retrieval” on OpenResearch

About this paper

Kai Zhang, Xinyuan Zhang, Ejaz Ahmed, Hongda Jiang, Caleb Kumar, Kai Sun, Zhaojiang Lin, Sanat Sharma and 6 more

AssoMem is a framework that enhances memory-augmented AI assistants for question answering by retrieving information from large personal memory repositories. It utilizes an associative memory graph and a multi-signal ranking system, RITRanker, which integrates relevance, importance, and temporal alignment. This approach yields improved accuracy and scalability in challenging similarity-dense memory environments, outperforming existing baselines.

Key insights

• Effective memory retrieval for AI assistants requires integrating multi-dimensional signals (relevance, importance, temporal alignment) rather than solely relying on semantic similarity, inspired by human associative memory.
• Graph-based memory structures, especially when combined with Personalized PageRank for importance scoring, provide a robust mechanism for organizing and retrieving context-aware information from large, dynamic memory banks.
• An adaptive fusion mechanism that dynamically weighs retrieval signals based on query intent is critical for achieving superior performance across diverse question types.

Summary by alphaXiv.