Lately, when developing with multiple CLI agents like Claude Code running at the same time, they often don't know what the others are doing, which leads to redundant waste of context and tokens. They're calling this the 'agent island tax.' It's interesting to see new architectures emerging to solve this.

One is an open-source project called kioku-mesh, which provides a shared memory layer using local SQLite or a Zenoh-based P2P protocol, allowing agents to share context and past decisions even while offline. The other is Cursor's Origin, released just a few days ago, which attempts to overcome latency and API limits by integrating agent memory graphs and execution flows directly into the Git hosting layer.

Instead of heavy synchronization via Redis or a central DB, this shift toward local-friendly P2P or tight integration with the editor's hosting infrastructure seems like it will become the trend. I'd love to test it out in my next toy project!