A teaching replay on synthetic data with a scripted agent. It demonstrates the
mechanics of memory poisoning and repair from published research; it is not evidence about the
safety of any real LLM or production memory system. Every number on this page is regenerated from
the repo's test suite.