Alibaba Releases MEMENTO: Qwen3-8B Context Agent Beats GPT-5.5

Alibaba researchers have unveiled MEMENTO, an open long-context agent built on Qwen3-8B that uses intelligent context management and note compaction to process up to 2.5 million tokens. By maintaining a strict 20,000-token context window, MEMENTO achieves higher accuracy than GPT-5.5 at a fraction of the inference cost.

What Is MEMENTO and How Does Alibaba's Context Management Work?

As large language models operate over increasingly vast inputs such as books, codebase repositories, and historical logs, keeping raw text inside an expanding attention window becomes computationally unsustainable. Traditional long-context execution relies on context compaction, where an agent summarizes previous conversation history into self-generated notes. However, off-the-shelf foundation models—including state-of-the-art architectures like GPT-5.5 and GLM-5.3—frequently generate vague or uninformative notes, leading to context decay over extended runs.

MEMENTO solves this fundamental limitation through a multi-stage training pipeline built atop Alibaba's open-weights Qwen3-8B model. First, the researchers bootstrap note-taking abilities by fine-tuning the base model on synthetic execution trajectories annotated with hindsight notes—summaries generated after the target answer is already known. Next, MEMENTO undergoes reinforcement learning (RL) guided by a composite reward function. This reward evaluates final answer correctness, note-quality heuristics, and model confidence calibration to ensure the agent retains critical details without hallucinating.

Despite being trained strictly to compact notes into a clean, fixed-size 20,000-token context window, MEMENTO naturally generalizes at inference time to advanced context strategies. These include context folding and confidence-driven backtracking, allowing the agent to revisit previous steps when uncertainty spikes during long-context processing.

How Does MEMENTO Outperform GPT-5.5 in Long-Context Tasks?

In extensive evaluations across four book-level question-answering benchmarks containing texts up to 2.5 million tokens long, MEMENTO achieved dramatic improvements over baseline models. Built on Qwen3-8B, the base architecture originally scored 36.9% accuracy when processing long-context books. With MEMENTO's context management framework enabled, Qwen3-8B's accuracy jumped to 67.8%, while never exceeding a 20K token context overhead.

When benchmarked against top proprietary competition, MEMENTO demonstrated superior cost efficiency and performance:

• Beating GPT-5.5 Agent: Under identical agentic evaluation harnesses, MEMENTO (67.8%) far outperformed GPT-5.5 in agentic mode (13.1%) as well as GPT-5.5 fed with full book context (56.7%), operating at less than 1/50th of the inference cost.
• Rivaling 100x Larger Models: MEMENTO nearly matched the performance of GLM-5.3 agent (69.3%) while operating at 1/9th the operational cost, and came within 6.4 percentage points of GLM-5.3 using full uncompacted context at 1/26th of the token expense.

These results establish that structured memory compaction in smaller models can outperform raw brute-force window expansion in frontier models.

Why Is Context Compaction More Effective Than Massive Token Windows?

Expanding context windows to millions of tokens introduces two primary challenges: exponential latency growth and context dilution (often referred to as the 'needle in a haystack' retrieval problem). When an LLM receives 2 million raw tokens, attention mechanisms become diluted across noisy, irremediable content, reducing answer precision.

MEMENTO addresses context dilution by transforming long-context understanding into an active context management problem. By teaching the model to continuously compress background information into structured, high-density notes and calibrating its confidence before finalizing context shifts, MEMENTO keeps the active working context sharp and hyper-focused.

This training methodology proves that reinforcement learning combined with hindsight synthetic data allows small-scale open-weights models to outmaneuver proprietary frontier models on complex document reasoning tasks.