SwiftMem brings fast memory recall to AI agents with smart query indexing

Researchers have unveiled SwiftMem, a memory architecture designed to help large language model agents retrieve information quickly and accurately without sacrificing performance as context grows.

The challenge of long-term memory#

Language model agents need to maintain and access long-term context to function effectively in real-world interactions. Traditional memory systems often slow down as they scale—even when backed by efficient vector search indexes like HNSW, they search across the entire memory space regardless of what the agent actually needs. This creates latency problems that make real-time agent interactions impractical.

How SwiftMem works#

The system narrows down memory searches to only the relevant subset of information through dual indexing across temporal and semantic dimensions. A temporal index enables quick time-sensitive lookups in logarithmic time, while a semantic index (called DAG-Tag) connects queries to relevant topics using hierarchical tag structures. To prevent the memory storage from fragmenting as it expands, SwiftMem introduces an embedding-tag co-consolidation mechanism that reorganizes information based on semantic clusters, improving how tightly related data is packed together.

Performance gains#

Testing across established benchmarks—LoCoMo and LongMemEval$_S$—shows SwiftMem achieving search latency of 10.8 to 12.7 milliseconds while maintaining competitive accuracy scores against HNSW-based memory systems. On a refined version of LoCoMo, the system preserves near-top-tier accuracy while delivering roughly a ten-fold latency advantage compared to existing approaches.

Learn more: View the paper on arXiv