INTERVENTIONS AS ADDRESSES: CAUSAL QUERIES AS ABSTAINING MEMORY READS

Estimators of causal structure whose parameters are shared across experimental conditions suffer from two structural vulnerabilities: catastrophic forgetting when experimental contexts arrive sequentially, and uncalibrated extrapolation when queried out of distribution or across counterfactual layers. We study an alternative, a memory of every episode stored under the experimental context that produced it, observational or which variable was set, with one rule added: it answers only where the stored evidence settles the query, and otherwise abstains. Held in a vector-symbolic memory with no trainable parameters, interventional distributions, conditioning and backdoor adjustment become reads and arithmetic. The abstention threshold is computed from the memory's own interference noise, so it needs no held-out data, and it is calibrated within a band of dimension. When experimental contexts arrive one after another, a gradient estimator forgets the earliest and the memory does not, since a sum does not depend on the order of its terms. An abstention comes for one of three reasons. A key, the address an episode is stored under, seen with two outcomes omits the variable that decides between them, and a search over the observables finds it. On Baba Is AI the search returns the rule text, and the memory then answers 0.977 of the held-out wall collisions it commits to while the rule is broken. On Alchemy it returns the displayed reward, and an agent given only the description language and the observation format reaches 0.91 of the ideal observer on held-out seeds. A key whose evidence several hypotheses still fit calls for an experiment, and expected free energy over a graded belief held beside the memory chooses it. A key naming two contexts at once can never be written, and the gate abstains from 60/60 such cross-world reads unprompted. Storing experience under its context thus gives an agent an interventional model that survives the order of its experiments, needs no retraining when a new variable is recorded, and says which of three things to do next when it cannot answer. This suggests that for an agent that learns by acting the useful form of causal knowledge may be a memory rather than a fit.