Why LLM Memory in Production Fails Silently

Published benchmarks show memory systems dropping to 48.6 at 10M tokens. The real fix is asserting on what retrieval returned before the model sees it.