Researchers have created a benchmark called MemTrapBench to test how large language models use memory. They found that even when memories are accurate and relevant, they can still distort model reasoning and affect performance on current tasks. This is known as a 'cognitive trap'. The authors propose a method called AdaptiveMem to mitigate these traps by instructing models to avoid them. Experiments showed that all evaluated memory strategies underperformed without memory, with some methods dropping over 10% in performance. --- Why it matters: This matters because it highlights the limitations of current large language model architectures and their reliance on memory. It also provides a new tool for researchers to evaluate and improve memory-based models. Source: https://arxiv.org/abs/2608.20202