MEMENTO: Learning Context Management for Long Context Processing Agents

MEMENTO is an agent that learns to manage its own context window by writing notes, and the paper's abstract reports that this lets an 8-billion-parameter model answer questions about very long books far more accurately, at a fraction of the cost of much larger models (arXiv). The technique is compaction: when the window fills, the agent replaces its history with a self-written note and keeps reading in a fresh window. The authors say off-the-shelf agentic models including GPT-5.5 and GLM-5.3 often write ineffective notes, and that reinforcement learning on its own rarely discovers good note-taking.

Their fix has two stages. They first fine-tune Qwen's Qwen3-8B on synthetic trajectories whose notes were written in hindsight, so the model sees what a useful note looks like, then train further with reinforcement learning using a reward that combines answer correctness, note-quality heuristics and confidence calibration. According to the abstract, on four book-level question-answering benchmarks with books up to 2.5 million tokens, accuracy rose from 36.9% to 67.8% while the agent never held more than 20,000 tokens in context. The authors report it beat a GPT-5.5 agent in the same harness (13.1%) and full-book context (56.7%) at under 1/50 the cost, and nearly matched a roughly 100 times larger GLM-5.3 agent (69.3%) at 1/9 the cost. It came within 6.4 points of GLM-5.3 reading the full book, at 1/26 the cost.

These are the authors' numbers from the abstract, on book question answering specifically. Other recent papers measure what compaction loses, such as Compact-Bench for coding agents and Lost in Compaction on user instructions lost under compaction. MEMENTO takes the other side of the problem and tries to train a model to write better notes.