Structured Distillation for Personalized Agent Memory: 11x Token Compression With 96% Retrieval Preservation via Compound Object Architecture

arXiv·medium signal

Paper proposes compressing conversation histories into a four-field compound object structure rather than raw text, achieving 11x token reduction while maintaining 96% retrieval quality. The compound object approach preserves semantic structure that flat compression destroys, enabling long-context agents to maintain effective memory at a fraction of the token cost. Directly actionable for builder deploying agents with persistent memory at scale.