A new arXiv paper introduces MEMENTO, an AI agent trained to decide what information to keep when working through very long texts. In tests on questions about books, the system raised Qwen3-8B’s accuracy from 36.9% to 67.8%, while using no more than 20,000 tokens of context at a time. The books in the tests were as long as 2.5 million tokens.
The approach tackles a basic problem for AI systems: they cannot always keep an entire long conversation or document in their working context. One common workaround, called compaction, is to replace earlier material with a shorter note. The paper says off-the-shelf agents, including GPT-5.5 and GLM-5.3, often make poor notes, and that reinforcement learning alone rarely teaches them effective note-taking.
How MEMENTO manages context
MEMENTO is built on Qwen3-8B. Its training began with synthetic examples in which notes were generated with hindsight, then used those examples to fine-tune the model. The team added reinforcement learning, a method that improves a model through feedback, with a reward based on answer correctness, note quality and how well the agent’s confidence matches its performance.
The system was trained to compress its history into a fresh context window. At test time, however, the paper says it also used approaches such as folding information back into its context and revisiting earlier material when confidence was low. These strategies help the agent work through lengthy books without loading the whole text into its active context.
What the results show
Across four book-question-answering benchmarks, the paper reports that MEMENTO scored 67.8%, compared with 13.1% for a GPT-5.5 agent in the same testing setup and 56.7% for a full-book-context comparison. It also scored 69.3% against a GLM-5.3 agent described as roughly 100 times larger. The paper reports that MEMENTO cost less than one-fiftieth as much as the GPT-5.5 agent and one-ninth as much as the GLM-5.3 agent; it came within 6.4 percentage points of GLM-5.3 using the full book context, at one-twenty-sixth the cost.
These are results reported by the paper’s authors, not evidence that MEMENTO will perform the same way on other tasks or documents. The abstract does not give the benchmark names, dollar costs or details about public access to the model. Its findings are specific to book-level question answering and the testing setup described in the paper.
Why it matters
For people building AI tools that need to answer questions about long books, MEMENTO suggests that learned note-taking can reduce how much text a model must hold at once. The reported gains are limited to book-question tests, so it remains unclear whether the method helps with other long documents or everyday uses.
