ConvMem: A Training-Free Hierarchical Convolutional Framework for Long-Context Reasoning

Sequential memory models for long-context reasoning process text chunk by chunk, often accumulating errors and suffering from high latency. ConvMem reframes this process as a hierarchical convolution: prompting an LLM serves as a convolutional kernel that compresses a linear sequence of segments into a logarithmic reasoning tree. By combining configurable strides, skip connections, and multi-kernel query decomposition, the framework achieves massive parallelism across segments without requiring reinforcement learning, maintaining robustness on out-of-distribution multi-hop benchmarks.