<![CDATA[Long-lived language-model assistants must decide what to store, which fact version to retrieve, when old evidence remains relevant, and when to reject an unsupported premise. This study evaluates CG-TRF, a confidence-gated temporal retrieval and forgetting policy, on the final ten-conversation LoCoMo benchmark: 272 sessions, 5,882 turns, 2,541 atomic observations, and 1,986 multi-hop, temporal, open-domain, single-hop, or adversarial questions. A fixed CPU-only extractive probe compared full history, a 120-turn window, session summaries, word TF-IDF, word-character hybrid memory, temporal-decay retrieval, and CG-TRF. CG-TRF combined confidence-gated writes, temporal and speaker alignment, revision links, 70% retention, diverse top-five retrieval, and confidence-based premise rejection. On 1,540 answer-bearing questions, it attained 0.1375 normalized token F1 and 0.0227 exact match, versus 0.1364 and 0.0175 for unrestricted hybrid memory. That F1 difference was not statistically supported, although CG-TRF stored 30% fewer records. It rejected 50.22% of 446 adversarial questions and reduced tempting-answer overlap to 0.0381; nonselective baselines rejected 0%. Its primary-budget Recall@5 was 0.4275 versus 0.5047 for hybrid memory; without retention, Recall@5 reached 0.5163 and F1 reached 0.1463. Confidence and attribution improved memory-layer safeguards, but forgetting caused the principal retrieval loss.]]>
Copyrights © 2026
