ConvMem: Convolutional Memory for Long-Context Reasoning
ConvMem replaces sequential long-context memory updates with a parallel hierarchy of summaries, offering a training-free pattern for agents that must reason over oversized inputs.
ConvMem treats a query-prompted LLM as a **convolutional kernel** that summarizes text segments through a hierarchical tree. This changes the reasoning path from a linear chain to a **logarithmic tree** and parallelizes work across segments and reasoning threads.
For agents processing large repositories or document sets, consider hierarchical, query-conditioned summaries instead of one sequential memory state. **Configurable Strides**, skip connections, and multiple semantic kernels are intended to retain evidence and separate complex query dimensions.
ConvMem treats a query-prompted LLM as a **convolutional kernel** that summarizes text segments through a hierarchical tree. This changes the reasoning path from a linear chain to a **logarithmic tree** and parallelizes work across segments and reasoning threads. For agents processing large repositories or document sets, consider hierarchical, query-conditioned summaries instead of one sequential memory state. **Configurable Strides**, skip connections, and multiple semantic kernels are intended to retain evidence and separate complex query dimensions. Results are limited here to claims on RULER-HotpotQA and RULER-2WikiMultiHopQA, with no scores, cost figures, or repository-scale coding evaluation supplied. Parallel calls may also introduce practical expense that the abstract does not quantify.
This adds hierarchical, parallel consolidation as a distinct long-context strategy: query-conditioned segment summaries can replace a single sequential memory path. It strengthens the case for preserving structure during retrieval, but introduces an unresolved systems tradeoff because logarithmic reasoning depth may require many parallel model calls and the supplied results do not quantify cost or coding-agent value.