SodaMem: Evidence-Grounded Temporal Graph Memory Enhances LLM Agent Recall

Researchers have introduced SodaMem, a novel memory architecture for large language model (LLM) agents designed to handle long-term conversational contexts. The system addresses a critical limitation in current LLM agents: their tendency to recall what was said rather than what is currently true. Traditional flat RAG diaries and Markdown logs, while efficient at retrieving specific information, fail to capture the currency, provenance, and ordered temporal reasoning necessary for sustained interactions. SodaMem tackles this by extracting typed FactEvents with mandatory provenance spans, persisting mention time, occurrence time, and validity, and linking them via SUPERSEDES, CONTRADICTS, and UPDATES edges under a hybrid lexical-dense indexing scheme. The architecture employs a planner-reader loop that gathers citable evidence before composing final responses, ensuring answers are grounded in verifiable facts. In evaluations on the LongMemEval-S benchmark, the store-of-record configuration achieved 92.8% accuracy (464 out of 500 questions, best of N=3) at a mean cost of $0.00161 per question (approximately 18.3k tokens; median $0.00111, approximately 14.6k tokens) using the deepseek-v4-flash model. This work, detailed in the arXiv paper 2608.08055, marks a significant step toward more reliable and context-aware AI assistants.