Long-term memory summaries retain only 3.05% of temporal expressions; a one-sentence prompt improves temporal QA accuracy by 31.4 points

Long-running agents often compress older conversations into summaries to control context length. New research suggests that the problem is not simply that “less information” remains, but that different types of information are removed at disproportionately different rates. The researcher used Salience-Weighted Consolidation (SWC) as a diagnostic framework: conversation segments are scored according to similarity to subsequent content, recency, and information density. High-priority segments are retained verbatim, medium-priority segments are rewritten as summaries, and low-priority segments are discarded.

The experiment used ten multi-session conversations from LoCoMo and compared truncation, rolling summarization, standard SWC, and temporally protected SWC under a 4,000-token budget. The main analysis covered 1,501 non-adversarial text questions. Claude Sonnet 4.6 served as the answer model, while Claude Haiku handled compression and evaluation. Standard SWC achieved an overall judged accuracy of 37.9%, outperforming rolling summarization at 23.8% and truncation at 17.1%. This indicates that structured consolidation can indeed preserve relationships between people and chains of events.

However, a clear gap emerged when the results were broken down by question type. Standard SWC scored 40.7% on multi-hop questions and 45.9% on single-hop questions, but only 15.6% on temporal questions. A retention-rate analysis identified the direct cause: of 952 temporal expressions in the original text, only 3.05% appeared verbatim in the summaries. By comparison, the retention rates for named entities and events were 8.03% and 5.03%, respectively. The summaries knew “who did what” but often lost “when” and “for how long.”

The researcher added just one sentence to the prompt, instructing the model to preserve dates, times, durations, ages, and relative-time expressions verbatim. This raised the temporal-expression retention rate to 62.39% and improved temporal-question accuracy by 0.314, from 15.6% to 47.0%. All ten conversations showed improvement in the same direction, while retention rates for people and events remained nearly unchanged. The implication for agent memory engineering is straightforward: compression specifications should explicitly identify protected data types, and evaluations should be stratified by temporal information, entities, commitments, and causal relationships.

The results remain a small-scale diagnostic study. It was conducted by a single researcher and tested only one SWC pipeline and one set of Claude models. Verbatim retention rates may fail to count semantically equivalent paraphrases, and the full-context upper bound was measured on only two conversations. Future work should determine whether the same fix transfers to rolling summarization, vector memory, and other models, and whether preserving more temporal information crowds out equally important permissions, numerical values, or commitment states.