Paper flags a scoring gap in AI memory retrieval

A memory can match your query perfectly and still fail to transfer, per a paper posted to arXiv. Most retrieval systems rank stored lessons on that resemblance alone, letting a close-looking match beat a more useful lesson to the top of the list.

Match score measures how close a stored lesson's wording looks to the current query. Transfer score, as the paper frames it, measures whether the lesson's underlying approach still works once the context changes. The paper says most systems only compute the first one, then rank and inject on that score alone.

The paper's fix: score every stored lesson twice, once for match and once for transfer, then rank on both before anything gets injected into context. A lesson can read as nearly identical to the query and still lack the knowledge that made it work the first time. Ranking on match alone can't tell the two apart.