The research paper A Survey on Large Language Models for Mathematical Reasoning reveals that while today’s large language models are excellent at reasoning in the short term, they are surprisingly poor at remembering information over time, a limitation that directly impacts how reliably they can serve as digital assistants or enterprise knowledge systems. Developed by researchers at Shanghai AI Lab and Fudan University, the study introduces MEMBench, the first standardized benchmark for measuring long-term memory in AI systems. It shows that even leading models like GPT-4o and Claude 3.5 lose up to 40% of their recall accuracy across extended sessions. For business and government leaders, this research matters because it exposes a key weakness in AI trust and continuity: without robust memory, AI cannot sustain contextual understanding, learn from ongoing user interactions, or manage long-term enterprise knowledge. The findings mark a crucial step toward developing persistent, personalized, and reliable AI systems that can truly integrate into daily operations.

_____

Key point: This paper introduces MEMBench, the first benchmark to measure how well AI models retain information over time, revealing that even top-performing systems suffer significant long-term memory loss, highlighting a critical barrier to building truly reliable, persistent AI assistants.