Beyond PLM argues that preserving product assumptions enables reassessment

AI-generated summary

Beyond PLM author Oleg argues that organizations must preserve the assumptions, constraints and rejected alternatives behind product decisions if they want to reassess those decisions later. Responding to Lionel Grealou’s article on whether enterprises remember what really matters, he agrees that historical evidence should not automatically become current guidance. Their disagreement concerns when to filter information: identifying important business questions today, Oleg argues, does not reliably identify all the context that will matter tomorrow. He defines Product Memory as a persistent, connected, explainable, attributed and recoverable layer linking product data to decision rationale. Both writers position Product Memory beyond any single enterprise application. PLM contains product definitions, requirements, configurations and engineering changes; ERP holds supplier, cost and procurement information; MES records shop-floor activity; and quality, regulatory and service systems contribute evidence. A material substitution illustrates the fragmentation: engineering approves a specification in PLM, ERP identifies its supplier and cost, MES records the affected batches, and quality establishes conformity. Explaining why the substitution happened and whether it succeeded requires connecting these records, often with email discussions. Because production failures, supplier constraints and regulatory changes also drive decisions upstream, Oleg advocates a connected graph of context above existing systems rather than another centralized vault or PLM module. His principal objection to selective preservation is that it repeats a weakness of traditional knowledge management: asking employees to predict future information needs and document them separately from their everyday work. Consider a team that rejects a supplier because of lead time. Three years later, an AI agent might recommend that supplier for its attractive pricing. Preserving the original rejection should not permanently disqualify the supplier; it should prompt an investigation into whether the lead-time constraint has changed. Without that context, the enterprise must repeat its evaluation or proceed without knowing what earlier work established. The proposed approach is broad capture with selective activation. Design-review comments, change discussions, supplier emails and conditional approvals should remain connected to the product objects they concern as a byproduct of work. This does not mean retaining everything indefinitely: access, confidentiality and retention boundaries still apply. Nor can any architecture recover assumptions that nobody expressed or recorded. At the point of use, governance should determine relevance to the question, configuration and lifecycle stage, source authority, and whether information has been superseded. Oleg acknowledges that an aerospace OEM and a consumer goods company will make different judgments. Traceability is essential to this approach but does not establish truth by itself. It records what happened, when and under whose authority, while Product Memory adds the reasoning. Trust also requires supporting evidence, applicability to the current situation and, where safety or compliance demands it, validation. Oleg endorses Grealou’s warning that AI-discovered relationships are not verified traceability. An agent linking a supplier email to an engineering change has generated a hypothesis, not an authoritative record. Systems of record retain their authority, and remembered context must distinguish their information from human statements and agent inferences. Addressing Grealou’s concern about obsolete reasoning and Jos Voskuil’s question about how knowledge graphs can “unlearn,” Oleg argues that reassessment depends on recoverable history. Earlier costs, regulations, supplier conditions and rejected alternatives provide the basis for testing whether a decision remains appropriate. Unlearning therefore means marking knowledge as superseded, not deleting evidence or rewriting history. Corrections should be recorded rather than overwrite prior context, preserving provenance while clearly separating historical evidence from current guidance. Grealou places memory within a broader “enterprise brain” and describes a cognitive data thread incorporating rationale and human judgment. Oleg accepts that memory alone is not intelligence, but argues that missing context contributes to AI errors such as choosing incorrect revisions, recommending unapproved substitutes or repeating rejected designs. His conclusion is to capture working context, trace required evidence and activate relevant knowledge without filtering away the basis for future challenges. The article presents a conceptual argument rather than measured implementation results. Oleg discloses that he is OpenBOM’s co-founder and CEO and that this work shapes his perspective.