Developer moves audit history out of prompts into a memory system for traceable drafts

A developer building an audit-drafting model found that stuffing past reports and findings into prompts was costly, mixed outdated and corrected facts, and could not be reproduced. They moved historical records into Hindsight, an open-source agent memory system, storing them as dated events. The prompt now holds only rules, rubric and new evidence, while past facts are retrieved on demand using combined semantic, keyword, graph and temporal search. Logging what was recalled alongside each draft created an audit trail that the earlier approach lacked. The author concludes that prompts act as working memory rather than storage, and that retrieval should be a visible, testable step.