MemStrata: 95% and 90.91% Source-Aware Accuracy on LongMemEval-500 and LoCoMo-1540 with a Local Qwen 3.8 27B Q4_K_M Reader

Researchers have introduced MemStrata, a system designed to improve how AI models generate answers by grounding them in full source documents rather than limited benchmark references. By testing against LongMemEval-500 and LoCoMo-1540 using a Qwen 3.8 27B Q4_K_M model, the study reports source-aware accuracy rates of 95% and 90.91%. This approach aims to address discrepancies between conversational outputs and incomplete reference data.