MemoryBench: Benchmarking Continual Learning in LLMs

Researchers have introduced MemoryBench, a benchmark designed to evaluate memory and continual learning capabilities in large language model systems (LLMsys). Recognizing that scaling data, parameters, and test-time computation is reaching diminishing returns due to high-quality data depletion and marginal gains, the field is shifting toward frameworks inspired by human learning and traditional AI systems. Existing benchmarks focus on homogeneous reading comprehension with long inputs, failing to test learning from accumulated user feedback. MemoryBench addresses this gap with a user feedback simulation framework and a comprehensive benchmark spanning multiple domains, languages, and task types. The work is published on arXiv under identifier 2510.17281.