{"slug": "memento-learning-context-management-for-long-context-processing-agents", "title": "MEMENTO: Learning Context Management for Long Context Processing Agents", "summary": "A Qwen3-8B-based long-context agent named MEMENTO, trained to manage its own context through compaction, raised book-level question-answering accuracy from 36.9% to 67.8% while never holding more than 20K tokens in context, according to an arXiv paper (2611.00244v1). On four benchmarks with books of up to 2.5M tokens, MEMENTO beat a GPT-5.5 agent in the same harness (13.1%) and full book context (56.7%) at less than 1/50 the cost, and nearly matched a roughly 100x larger GLM-5.3 agent (69.3%) at 1/9 the cost. The method bootstraps note-taking by fine-tuning on synthetic trajectories annotated with hindsight-generated notes, then reinforces it with RL using a reward combining answer correctness, note-quality heuristics and confidence calibration.", "body_md": "arXiv:2611.00244v1 Announce Type: new \nAbstract: LLM agents that operate over long inputs must perform context management: deciding what to retain in a fixed-size context window. A common approach is compaction, where the model replaces its history with a self-generated note. Yet off-the-shelf agentic models, including GPT-5.5 and GLM-5.3, often write ineffective notes, and reinforcement learning (RL) alone rarely discovers useful note-taking behavior. We present MEMENTO, a long-context agent built on Qwen3-8B and trained to manage its context through compaction. We first bootstrap note-taking by fine-tuning on synthetic trajectories annotated with notes generated in hindsight, then reinforce it with RL using a reward that combines answer correctness, note-quality heuristics, and confidence calibration. Although trained only to compact into a fresh window, MEMENTO extends at test time to strategies such as context folding and confidence-based backtracking. On book-level question answering across four benchmarks, with books of up to 2.5M tokens, MEMENTO raises Qwen3-8B's accuracy from 36.9% to 67.8% while never holding more than 20K tokens in context. It outperforms GPT-5.5 agent in the same harness (13.1%) and full book context (56.7%) at less than 1/50 the cost. It also nearly matches a ~100x larger GLM-5.3 agent (69.3%) at 1/9 the cost, and comes within 6.4 points of GLM-5.3 full context at 1/26 the cost.", "url": "https://wpnews.pro/news/memento-learning-context-management-for-long-context-processing-agents", "canonical_source": "https://arxiv.org/abs/2611.00244", "published_at": "2026-11-04 05:00:00+00:00", "updated_at": "2026-11-04 05:20:52.632965+00:00", "lang": "en", "topics": ["ai-agents", "large-language-models", "ai-research", "artificial-intelligence", "machine-learning"], "entities": ["MEMENTO", "Qwen3-8B", "GPT-5.5", "GLM-5.3", "arXiv"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/memento-learning-context-management-for-long-context-processing-agents", "markdown": "https://wpnews.pro/news/memento-learning-context-management-for-long-context-processing-agents.md", "text": "https://wpnews.pro/news/memento-learning-context-management-for-long-context-processing-agents.txt", "jsonld": "https://wpnews.pro/news/memento-learning-context-management-for-long-context-processing-agents.jsonld"}}