Preprint proposes zero-inference prospective term to improve personal memory retrieval

A new arXiv preprint introduces a zero-inference 'prospective term' designed to improve personal memory retrieval by incorporating user commitments stored in an explicit ledger. Commitments are recorded as dated or trigger-conditioned entries, and memory items linked to a firing entry receive a salience boost that is blended multiplicatively into embedding-based retrieval, ensuring that relevance remains the primary ranking signal.

The authors evaluate the approach on a synthetic prospective-memory task set modeled after TriggerBench’s published structure, comprising 48 blind-authored dialogues and 175 tasks. On the hard stratum, the prospective term raised recall@5 from 0.000 to 0.955 at the default blend weight and to 1.000 under a floor variant, with zero false boosts observed across 53 resolved-commitment tasks.

The authors also report that only 17–29% of naturally phrased commitment-trigger pairs defeat embedding similarity, indicating the prospective term is impactful on a minority of cases and does not harm performance on the remainder. They position precomputed commitment linkage as the always-on floor of a layered design, with query-time prospection serving as an expansion layer.

The results are described as preliminary, with the evaluation set constructed by the authors and plans to evaluate on TriggerBench proper once its data is released. The preprint is seven pages and includes links to code and data.