This paper measures the real impact of adding a persistent memory tier to multi‑agent LLM inference. Decomposing long contexts across cooperating agents limits the KV‑cache working set to 14.3 MiB per query, compared with 35.5 MiB and 35.3 MiB for single‑pass and retrieval‑augmented baselines. The persistent tier stores reasoning traces, yet on eight controlled dataset pairs (100 samples each) it adds about 0.368 MiB of peak cache (+0.167, +0.590) and yields no statistically significant accuracy gain (+0.015, 95% CI [-0.011, +0.046]). The authors argue the null result is structural: single‑question benchmarks supply each item with its own evidence and score it independently, requiring stored traces to be reset between conditions, so recall has nothing informative to retrieve. Four measurement corrections were needed—three inflating apparent benefit, the fourth making a resolvable effect—none of which appear in the results table. The paper outlines the conditions an agent‑memory ablation must satisfy and detection procedures that do not rely on knowledge of the specific defect.
Review