SwiftMem: Query-Aware Agentic Memory System for LLMs

A new research paper introduces SwiftMem, an agentic memory system designed to reduce latency in large language model (LLM) agents by enabling query-aware retrieval. Unlike existing frameworks that perform full-scope searches over the entire memory embedding space, SwiftMem narrows retrieval to query-relevant subsets using specialized indexing over temporal and semantic dimensions. The system includes a temporal index for logarithmic-time range queries and a semantic DAG-Tag index that maps queries to relevant topics via hierarchical tag structures. To manage memory growth, SwiftMem employs an embedding-tag co-consolidation mechanism that reorganizes memory to prevent fragmentation. The paper is available on arXiv under identifier 2601.08160.