Persistent Memory in Multi-Agent LLM Inference: What It Costs, What It Buys, and When You Can Tell

Decomposing long-context inference across cooperating agents bounds the key-value cache that each call holds, rather than the total evidence, which matters when cache memory is the binding limit. Many such systems add a persistent tier that stores and recalls reasoning traces, and an ablation that reports an accuracy gain usually supports that tier. We measure both on one three-tier agent architecture. Decomposition delivers a peak key-value working set of 14.3 MiB per query, against 35.5 and 35.3 MiB for the single-pass and retrieval-augmented baselines. The persistent tier does not. Across eight controlled dataset pairs at n=100 questions per arm, it costs +0.368 MiB [+0.167, +0.590] of peak cache and produces no detectable accuracy change (+0.015, 95% confidence interval [−0.011, +0.046]). We argue that this null result is structural. Single-question benchmarks supply each item with its own evidence and score it independently, and a correct measurement resets stored traces between conditions, so recall has nothing informative to retrieve. Reaching this result took four measurement corrections, and none was visible in the results table. Three inflated the apparent benefit, and the fourth made an effect of that size look resolvable. We give the conditions that an agent-memory ablation must satisfy, with detection procedures that need no knowledge of the specific defect.