C3M is a memory organization method for long-horizon tasks involving multiple sessions and multimodal data (text and images). It keeps a compact active index of routing representations while preserving all original evidence in a persistent cold store, updating the index with relation-aware actions to avoid conflating similar but distinct records. At query time, it routes within a budget to retrieve relevant source evidence.
Experiments on MemLens show C3M outperforms existing compression-based baselines (ReSum, MovieChat, MemRefine, CDRs) on overall accuracy for both GPT-5.6 and Qwen 3.8 Flash at history lengths 32K and 64K, though it sometimes falls behind at 128K. The method is particularly strong at maintaining temporal distinctions and multimodal evidence, which is critical for tasks like state tracking and knowledge updating.