How a Single OCR Misread Can Corrupt an AI Agent's Long-Term Memory

A vision model misread a smudged diary entry, interpreting '18 June' as '13 June,' a small error that became a confidently cited legal fact once stored in an AI agent's long-term memory. The incident highlights how AI memory layers extract, link, and summarize stored data without questioning its accuracy, turning input errors into authoritative misinformation. The developer concluded that data must be validated before being committed to memory, not after, and that deterministic rule-based checks should only lower — never raise — a model's confidence score. Keeping original source files permanently attached to every memory entry was identified as a critical safety net, allowing users to verify recalled facts against raw documents. The author also recommended designing idempotent re-upload systems so corrected documents replace erroneous memories rather than creating conflicting duplicate entries.