Ramp Labs has introduced "Latent Briefing," a new method for efficient memory sharing in multi-agent systems, significantly reducing token usage by up to 65% without compromising accuracy. The approach compresses large model KV caches, allowing for more efficient task decomposition and execution in multi-agent architectures. On the LongBench v2 benchmark, the method showed a 65% reduction in token consumption for worker models and improved accuracy by 3 percentage points. The solution, tested with Claude Sonnet 4 and Qwen3-14B models, demonstrated faster processing times and adaptability across various document types.