When Experience Survives the Agent
Testing Causal Operational Inheritance Across Successor Replacement
When does persistence become causal inheritance?
Long-lived artificial agents can preserve facts, trajectories, procedures and other forms of memory across sessions and model boundaries. Persistence alone, however, does not establish that a predecessor's experience remains causally operative in a successor.
This paper introduces causal operational inheritance (COI) as an interventionally testable predecessor–successor relation: the inherited object must be operational, its formation must depend on predecessor experience, its effect must survive removal of that experience and replacement of the active carrier, and the proposed mediator must withstand causal intervention.
A six-part evidentiary assay.
The proposed COI-6 criterion requires six steps: experience intervention, transportable-state mutation, origin removal, fresh-carrier instantiation, counterfactual successor divergence, and mediator intervention.
The positive result is deliberately narrow.
In PLI-2G, corrective experience produced a token-specific operational mapping. After the originating history was removed, fresh carriers followed that mapping across all three synthetic domains tested.
PLI-2G-M then targeted the proposed mediator. Mapping-only transplant reproduced the effect in 3/3 domains; mapping ablation removed it despite retained semantic residue in 3/3; and wrong-token sham mappings produced no effect in 3/3. Within this engineered representation, SELECTOR_MAP was necessary and sufficient for the tested token-specific effect.
Negative controls define the boundary.
PLI-2H did not pass as a discriminating test of less-prestructured significance selection: both H0 and ablation validity controls reproduced the H1 forced-choice pattern. The paper therefore withholds the stronger interpretation.
The proof-of-concept is local to one engineered updater and state representation, fresh sessions within one tested model environment, three synthetic domains and deterministic-style forced-choice tasks. It does not establish cross-model or cross-provider robustness, autonomous significance selection, general developmental continuity, consciousness, personhood or identity persistence.
Read the paper and inspect its evidentiary boundary.
The publication records both positive and failed experiments because the negative results define which stronger interpretations the present evidence does not support.
Replication should be a new test, not a repair.
The paper identifies two next directions: a prospectively frozen replication with repeated stochastic runs, an independently implemented updater and additional model/provider environments; and a separate significance-selection assay that manipulates the lineage-level criterion of relevance itself.
Criticism, replication designs and alternative implementations are welcome when they make the evidentiary claim more testable rather than broader by assertion.