Semantic Density Does Not Predict Agent-Memory Collapse

Persistent textual memory can improve LLM agents, but repeated updates can erode task performance. We ask whether memory-store statistics warn before held-out accuracy declines. We registered mean pairwise cosine similarity ($\rho$) and 13 additional geometry, retrieval, and graph statistics on 805 snapshots from 13 memory methods across ALFWorld, AppWorld, ScienceWorld, WebShop, and two ARC-AGI settings. These methods store trajectories, rules, workflows, strategies, or linked records, using selective retrieval or whole-store prompting. No such statistic reliably predicted early signs of task accuracy decline, and $\rho$ had zero median lead. With the checkpoint after peak accuracy as a secondary decline point, measures of highly similar pairs and nearest-neighbor crowding changed earlier but did not consistently outperform stored-item count. We trained classifiers on this corpus and applied them unchanged to PropWorld runs of up to 5,000 crafting tasks. They falsely flagged 56.9\% of checkpoints from size-limited stores with stable accuracy and detected only 18\% of checkpoints after task success fell to zero. Thus, global semantic density did not provide a reliable warning, while local crowding did not transfer consistently across memory designs. In these long-horizon runs, size-limited stores remained stable, while the sole failure arose when unbounded whole-store prompting exceeded model input capacity, which token growth predicted directly.