Original title: ConvMem: Convolutional Memory for Long-Context Reasoning
Hongming Zhang, Zhaozhen Gu, Fengshuo Bai, Ming Hao +5 · September 2026
Large Language Models often struggle with very long texts because they can only process a fixed amount of information at once. Past solutions read long documents segment by segment, updating a working memory as they go. This process is slow, accumulates errors, and often requires expensive extra training. To solve this, researchers built ConvMem, an approach that requires no new training. ConvMem treats the language model as a filter that summarizes small text pieces and stacks them into a tree-like pyramid, rather than a slow, single-file line. It processes multiple sections and complex query parts at the same time, while using shortcuts to keep key facts from getting lost. In tests, ConvMem beat comparable training-free methods and avoided overfittingWhen a model learns a task too well on specific training data, making it perform poorly on new, unseen data. when facing unfamiliar tasks.
Imagine you're trying to find a specific detail in a huge library. Normally, you'd have to walk down every single aisle, shelf by shelf, reading every book title. This takes forever. Now, imagine you have a special assistant who can quickly glance at a whole section of shelves and give you a one-sentence summary of what's there. You can then use these summaries to quickly narrow down which sections to look at more closely. In this paper, the LLM acts like that assistant, summarizing chunks of text, and these summaries are organized in a tree so the LLM can find the information much faster.
Large language models can get lost in very long texts, making them unreliable for tasks like summarizing books or analyzing lengthy legal documents. This new method allows them to process more information efficiently and accurately, making them more useful for complex real-world applications.
The abstract does not specify limitations beyond stating that the method outperforms training-free baselines and avoids overfitting seen in RL-trained models on out-of-distribution tasks. It does not mention performance on specific types of long-context tasks or potential computational costs.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
ConvMem: Convolutional Memory for Long-Context Reasoning — Hongming Zhang, Zhaozhen Gu, Fengshuo Bai et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.