Efficient Memory Sharing for Multi-Agent Systems via KV Cache Compaction