GRAVITY: Structured Memory Module for Long-Horizon Conversational Agents

Researchers have unveiled GRAVITY (Generation-time Relational Anchoring Via Injected Topological Memory), a modular memory system that can be easily integrated to improve long-term conversational agents. In contrast to conventional retrieval methods that provide unstructured text to language models, GRAVITY derives three distinct knowledge representations from raw dialogue: entity profiles based on relational graphs, causal traces formed from temporal event tuples, and summaries of topics across sessions. During the generation phase, these representations are incorporated into the host system's prompt as structured anchoring contexts, allowing the model to compile fragmented evidence into coherent, relevant responses. This method tackles the reasoning deficiencies present in current memory systems lacking relational, temporal, and thematic frameworks. The research was published on arXiv under identifier 2605.01688v1.