VikingMem: A Memory Base Management System for Stateful LLM-based Applications

Large Language Models have transformed interactive applications, but their limited context windows create challenges in managing long-term interactions. Existing memory management methods often rely on simplistic extraction techniques, leading to incomplete memories or are too rigid for diverse applications.

VikingMem addresses these issues by introducing a Memory Base that emphasizes selective extraction of valuable memories, stateful evolution of memory content, and a generalizable abstraction for various applications. Implemented on the VikingDB vector engine, VikingMem utilizes event-centric memory extraction and dynamic updates to maintain statefulness.

It employs temporal compression through a topic-wise timeline, prioritizing recent interactions while summarizing and fading older memories. Extensive evaluations on long-term memory benchmarks indicate that VikingMem outperforms existing methods by up to 30% in memory retrieval effectiveness, crucial for maintaining low latency in interactive settings.

This advancement opens new avenues for developing more effective stateful applications across multiple domains.