Cosine Similarity Said 0.92. The Payload Was Already Dead.

The author explores the critical gap between an LLM agent's memory retrieval and actual behavioral change, challenging the common assumption that persistent memory automatically leads to 'poisoning.' By implementing a rigorous measurement rig based on counterfactual amputation, the study demonstrates that even when embedding cosine similarity scores are high (e.g., 0.92), the payload may be functionally inert. The article argues that replication should focus on the preservation of causal behavior rather than mere representational similarity. Through various experiments with local models, the author shows that many perceived 'AI worm' scenarios fail to produce measurable behavioral shifts when properly instrumented. The post provides actionable practices for developers, including logging retrieval as an event, tracking parse failures as first-class metrics, and running negative controls to establish a noise floor. Ultimately, the author emphasizes that without causal instrumentation, high similarity scores can be misleading, masking the actual failure of a payload to influence agent actions.