An assistant can find a memory about the right topic and still retrieve the wrong event. A recurring dinner, meeting or journey may resemble dozens of others. TIDER addresses that problem by predicting a useful time window before choosing the evidence for an answer.
The paper argues that timestamps added during storage or reasoning are insufficient if the correct memory has already been displaced from the retrieval results. Its intervention therefore happens when candidates are selected, before the answering model receives its context.
Two routes to the evidence
TIDER uses a reinforcement-trained temporal locator alongside semantic retrieval. The locator takes a query and timeline summary, identifies whether time is relevant and predicts intervals likely to contain the answer's evidence. A temporal channel retrieves within those intervals; its candidates are merged with the semantic channel and reranked by proximity to the predicted windows.
The architecture keeps the two signals separate. A memory can qualify because its content is similar, because it falls in the predicted period, or through both routes. For non-temporal questions, the paper describes closing the temporal channel and falling back to semantic retrieval.
What the reported results establish
The authors use a 1.7-billion-parameter locator. They report accuracy of 61.52% on LifeBench and 78.96% on LoCoMo, with improvements of 8.82 and 3.66 percentage points respectively in their comparison. These are results from the paper's evaluation, not measurements of a Franklin assistant.
The training reward combines temporal need, window localisation, evidence coverage and answer correctness. That design makes the interval decision part of the learned retrieval policy rather than a fixed date filter applied to every question.
When this is useful to investigate
The most relevant application is a long-running assistant whose history contains many similar events. A useful evaluation would deliberately include near-duplicate memories from different dates, along with questions where no time restriction should be applied. That would expose whether a temporal improvement comes at the expense of ordinary recall.
Our interpretation is that retrieval quality should be inspected at the evidence level as well as through final answer accuracy. Record which interval was selected and whether the required memory entered the candidate set. A better-sounding answer cannot recover a missing event, and a mistaken interval can exclude the very evidence the temporal channel was intended to rescue.
Comments