MEMENTO agent improves long-context QA accuracy from 36.9% to 67.8% while limiting context to 20K tokens
A Qwen3-8B-based agent trained with synthetic note-taking and RL outperforms a GPT-5.5 agent by 54.7 percentage points and nearly matches a ~100x larger GLM-5.3 agent at a fraction of the cost.
1 source · cross-referenced
- MEMENTO, a long-context agent built on Qwen3-8B, raises accuracy on book-level QA from 36.9% to 67.8% while keeping context under 20K tokens.
Researchers from Columbia University and other institutions introduce MEMENTO, a long-context agent built on Qwen3-8B that learns context management through compaction. The approach combines fine-tuning on synthetic trajectories with notes generated in hindsight and reinforcement learning using a reward combining answer correctness, note-quality heuristics, and confidence calibration.
On book-level question answering across four benchmarks with books up to 2.5M tokens, MEMENTO improves Qwen3-8B’s accuracy from 36.9% to 67.8% while never holding more than 20K tokens in context. The agent also generalizes at test time to strategies such as context folding and confidence-based backtracking, even though it was trained only to compact into a fresh window.
In the same evaluation harness, MEMENTO outperforms a GPT-5.5 agent by 54.7 percentage points (67.8% vs. 13.1%) and exceeds a GPT-5.5 agent using full book context (56.7%) at less than 1/50 the cost. It 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 performance at 1/26 the cost.
- Nov 11, 2026 · Latent Space — swyx
OpenAI demonstrates voice steering, cloud handoffs, and personal agents for Codex at AI Engineer Code Summit
Trust79 - Nov 11, 2026 · The Verge — AI
OpenAI to launch verified AI Business Agents in ChatGPT for US e-commerce and service businesses
Trust79 - Nov 10, 2026 · Latent Space — swyx
NaiveAI releases Naive-N0.5-Flash coding model with ultrafast token throughput
Trust74