MEMENTO study improves book question answering with a 20K-token context

Researchers report in a preprint that MEMENTO raised Qwen3-8B's accuracy from 36.9% to 67.8% across four book question-answering benchmarks while holding no more than 20K tokens in context. The tests included books of up to 2.5M tokens. MEMENTO learns to replace its history with notes, first through training on synthetic examples and then through reinforcement learning that rewards correctness, note quality and confidence calibration. In these tests, it exceeded GPT-5.5's full-book-context accuracy of 56.7% at less than 1/50 the cost. GLM-5.3's agent scored higher, at 69.3%, while MEMENTO used 1/9 the cost.