Your organization already knows what causes what. The knowledge is just not where you can use it.

The knowledge problem

The most consequential decisions in an organization are causal. What will happen if we change this? What would have happened if we hadn’t? Why did this fail in case A but not case B? These are Rung 2 and Rung 3 questions in Pearl’s ladder of causation, and they require a causal model to answer.

The model is not the bottleneck. The bottleneck is the content: the structural beliefs, the parameter priors, the edge-case knowledge that the model needs to be built from. That content already exists — in the senior actuary’s understanding of the loss curve, in the clinician’s twenty years of patient outcomes, in the process engineer’s record of every failure mode the equipment has shown. The content has three properties that make it unusable in its current form:

• It is unwritten. The actuary has never been asked to articulate the full causal structure of the underwriting process. The clinician has never been asked to lay out the full DAG of treatment effects. The expert knows the structure; the structure has never left their head.
• It is unstructured. Even when it has been written — in white papers, post-mortems, internal documents — the writing is prose. Prose cannot be queried, composed, or validated. Two prose accounts can contradict each other without anyone noticing, because there is no structure within which a contradiction is visible.
• It is unprotected. The retirement of one senior expert removes a portion of the organization’s decision-making capability that is, in principle, irreplaceable. The knowledge took decades to accumulate. It exits in a week.

The pattern is consistent across industries. Insurance, healthcare, manufacturing, defence, finance, pharmaceuticals: the most important knowledge is in the fewest heads, written down in the fewest places, and protected by the weakest mechanisms.

What causal memory is

Causal memory is a structured, queryable representation of an organization’s beliefs about cause and effect — what affects what, under what conditions, with what strength. The representation is a set of directed acyclic graphs, with edges that name causal mechanisms and nodes that name the variables those mechanisms relate.

The representation is not a simplification of the expert’s knowledge. It is the part of the expert’s knowledge that is decision-relevant: the graph structure (what causes what), the strength of relationships (parameter priors), the conditions under which each relationship holds (scope), and the cases that contradict the simple version (edge-case exceptions). Everything decision-relevant has a place in the structure.

For the full position on what an organization’s causal knowledge looks like when it’s made permanent — what it contains, how it’s elicited, what it protects against, and how it differs from the documentation organizations currently produce — see An Organization’s Causal Knowledge.

Three properties of the representation matter for what it enables:

• It is composable. Two experts’ graphs can be joined — where they agree, the joint graph carries both perspectives; where they disagree, the disagreement is surfaced as an explicit structural question.
• It is queryable. A question stated in plain language — “what happens to loss ratio if we tighten the underwriting on this segment?” — can be translated into a formal causal query against the graph.
• It is auditable. Every edge in the graph can be traced to the elicitation transcript that introduced it. Every parameter prior can be traced to the expert who supplied it.

What it enables: Rung 1, 2, 3 operations

Once the causal knowledge is structured, it supports operations at each rung of Pearl’s ladder of causation. The same memory answers different kinds of questions depending on the operation applied:

• Rung 1 — Associational. What patterns are present in our historical data? Standard descriptive analytics. The causal memory is not strictly required for Rung 1 — ML on the historical data works — but the memory provides the variable definitions and the scope boundaries that keep the answers meaningful.
• Rung 2 — Interventional. What happens if we change X? Requires the causal graph plus the do-operator. The memory is the graph; the do-calculus is the operation. The answer is computed against the structure the experts laid down, not the historical correlations.
• Rung 3 — Counterfactual. What would have happened if we hadn’t changed X, in this specific case? Requires the structural causal model — the graph plus the exogenous noise terms that account for individual specificity. The memory is the graph and the parameter priors; the counterfactual reasoning is the operation.

The same causal memory supports all three. An organization that has its causal knowledge in this form has not committed to a particular kind of question — it has built infrastructure that supports any of the three when the question arrives.

Rung 3 counterfactuals as the canonical example

Rung 3 is the operation a causal memory enables that no other infrastructure does. Counterfactual reasoning — “what would have happened in this specific case under the alternative?” — is what regulatory audits, clinical-governance committees, adverse-event reviews, and adjudication boards actually require. None of these are answerable from historical data alone, no matter how much of it exists.

The procedure has three steps: abduction (inferring the individual-specific factors from what was observed), action (modifying the intervention variable), and prediction (re-running the model under the modified intervention with the inferred individual-specific factors held fixed). See Rung 3 — Counterfactual Reasoning — The Three-Step Procedure for the worked machinery.

The procedure is operationally accessible only when the causal memory is in place. Without the structural model and the parameter priors, abduction has nothing to infer from; without the graph, action has no variable to intervene on; without the structural equations, prediction has no mechanism to propagate the change. The Rung 3 operation is what the memory unlocks, and the memory is the prerequisite the operation can’t be performed without.

Surviving turnover, audit, and regulatory review

Once the causal knowledge is structured, the three failure modes of the unwritten version are closed:

• Turnover. The retirement of one expert no longer removes the knowledge. The structural beliefs are in the graph; the parameter priors are recorded with their elicitation source; the edge cases are documented with the cases that produced them. A successor inherits an asset, not a vacancy.
• Audit. A regulator or audit committee that asks “how did you arrive at this conclusion?” receives the trace from question to answer with every assumption inspectable. The model’s estimate is not the artefact — the assumption-trace is. The conversation moves from defending a number to defending a structure, which is the conversation the experts are equipped to have.
• Regulatory review. The structural model is the artefact that survives changes in the underlying ML implementation, the data warehouse, the analytical tooling. A regulator looking at the model in five years sees the same graph the experts built, modified only by reviewer feedback that is itself recorded. The structure is durable; the operational substrate around it can be replaced.

The Library architecture — The Library — is the system that holds the causal memory and makes these properties operationally available. Causal memory is the content; the Library is the structure that contains it.