Medication reconciliation is commonly represented as the production of one accurate medication list from several incomplete records. This representation obscures a more difficult problem: medication records are source-specific accounts produced for different purposes, at different times, and under different documentation conditions. Prescribing records, dispensing histories, administration records, clinical notes, and patient reports may therefore disagree without any source being uniformly authoritative. This article proposes a multi-agent medication-record reconciliation architecture that treats each record entry as a temporally situated assertion rather than an unquestioned fact. Bounded source-specific agents extract medication assertions while preserving provenance, documentation time, inferred validity intervals, uncertainty, and source limitations. Shared services normalize medication identity, classify disagreement, assess evidence suitability, and construct conflict sets without silently deleting competing assertions. A reconciliation orchestrator generates provisional medication states, while an independent challenge function examines unsupported assumptions and correlated agreement. Conflicts exceeding configured authority or evidentiary sufficiency are transferred to structured human adjudication, external verification, or safety escalation. The proposed architecture distinguishes semantic, temporal, intentional, provenance, omission, and uncertainty disagreements and separates technical reconciliation performance from workflow usefulness, medication safety, equity, and patient outcomes. Validation would require staged component testing, multisite retrospective assessment, silent prospective evaluation, human-factors studies, comparative workflow evaluation, subgroup analysis, clinical-outcome investigation, and postdeployment monitoring. The architecture is an original non-empirical synthesis. It does not establish that multi-agent reconciliation is superior to conventional software or pharmacist-led practice, and it should not be interpreted as a clinically validated, autonomous, or deployment-ready system.