Woofun AI reports that researchers from Tsinghua University, Shanghai Jiao Tong University, and MemTensor published a systematic evaluation of 12 mainstream large-scale intelligent agent memory systems, including Mem0, Letta, and Zep. The team introduced a four-module analytical framework covering memory representation, retrieval, routing, and maintenance, quantifying performance across 11 datasets.

The findings indicate that no single memory architecture adapts to all workloads. Hybrid systems excelled in conversational question-answering, while structured topology systems proved reliable for single-step fact recall but struggled with temporal reasoning. Purely appended memories suffered catastrophic degradation during long runs, with standard semantic consolidation disrupting temporal cues and causing "past hallucinations." Additionally, traditional similarity retrieval accuracy declined sharply over longer time spans, and highly structured graph systems incurred high latency without proportional accuracy gains.