Most AI agent 'memory' is actually a session transcript piped into a vector database, with similarity search substituting for true recall. This architectural shortcut fails audits, complicates vendor swaps, and loses the source interaction record once raw events age out. The argument here is that durable agent memory should be an immutable, ordered event log of every tool call, model output, and state transition — with retrieval layers (vector indices, summaries, knowledge graphs) derived from that record rather than replacing it. Four deployment patterns are outlined: sidecar, replace, complement a managed memory service, or greenfield. The practical stakes are illustrated with a fraud scenario where a regulator demands point-in-time reconstruction of a $4M wire transfer decision.
