When the retrieval cue is based on event content (what happened), models perform best; when the cue is spatial (where), performance drops; and when the cue is temporal (when), performance is worst. This gradient is visible across all six models in both in-context and RAG settings, and holds for both the short and long books. For GPT-4o in-context on the long book, single-event F1 is 0.93 for content cues, 0.93 for space cues, 0.97 for entity cues, but only 1.00 for time cues at zero events dropping to 0.47 at 6+ events. The authors note this mirrors human memory research where temporal memory is less reliable than spatial or content memory.
Evidence
correlational
Key metric
GPT-4o in-context (long book, 1 event): content 0.93, space 0.93, entity 0.97, time 1.00; (6+ events): content 0.56, space 0.50, entity 0.59, time 0.47
Caveat
The authors note it is unclear whether humans are better or worse at temporal versus spatial memory, and that the benchmark uses explicit date strings rather than implicit temporal references.