Agent memory v2 — seven rules after the poisoning

Short summary

Agent memory systems amplify hallucinations when re-ingesting model output as fact. The fix: add provenance tracking (verified/unverified/unavailable) to memory schema, audit logging on state changes, and demotion logic when contradictions surface. Seven architectural rules prevent silent hallucination storage.

• •Author's agent stored model hallucinations as facts by treating its own summaries as verified sources
• •Root cause: no provenance field to distinguish verified facts from model assertions
• •Solution: three-valued provenance states, immutable audit logs, and contradiction-based demotion of disputed claims

Generated with AI, which can make mistakes.