Agentic Memory for Long-Term Temporal Knowledge

Knowledge representation and retrieval for long-term temporal information in Agentic Memory systems presents unique challenges in maintaining context across extended time periods while ensuring efficient access to relevant historical data. While prior works have focused on conversational memory, long-term semantic memory remains an under-explored area. The core objective is to develop methodologies that efficiently and effectively represent entities and their relationships—particularly temporal relationships—for answering complex multi-step temporal queries.

This Summer internship project I mentored aligns with long-term memory initiatives within our organization, contributing to AWS's broader AI/ML research portfolio and advancing capabilities for enterprise AI applications.

• Advances understanding of temporal knowledge systems and their applications in enterprise environments, particularly for AI-powered document analysis and decision support systems.
• Enables real-world use cases such as financial analysts exploring document corpora where entity relationships evolve over time, providing temporal context for strategic decisions.
• Addresses the non-trivial challenge of constructing coherent chains of events that link different entities across temporal sequences, essential for complex analytical workflows.

• Dataset & Benchmark Development: Created a large-scale dataset and benchmark for complex multi-step temporal question answering involving multiple entities, establishing evaluation standards for temporal knowledge systems.
• Temporal Knowledge Graph Design: Developed an innovative memory representation utilizing a novel knowledge graph formulation that explicitly encodes temporal relationships between entities and events.
• Agentic RAG Architecture: Designed a novel retrieval and ranking architecture for Agentic RAG systems, optimizing performance for temporal query processing and knowledge retrieval.