Developer builds two-stage memory recall to prevent vendor data crowding in AI agents

A software developer published a technical walkthrough on DEV Community describing how they solved a retrieval problem in an AI procurement agent where high-volume vendors were pushing out relevant records from lower-volume ones. The solution uses a two-stage recall system: the first pass fetches only exact tag matches (both vendor and exception type), while the second pass broadens the search if the first returns too few results. This approach leverages two tag-matching modes — 'all_strict' and 'any_strict' — available in the Hindsight memory library to control result precision. Only records from the exact-match pass are fed to the language model for decision-making, while broader results are surfaced in the UI for transparency but cannot influence recommendations. A unit test confirms the fix works: a vendor with just one stored decision is no longer displaced by a competing vendor with four entries under a single-pass retrieval approach.