Concept Containers as Representation-Level Regulation in Artificial Agents

We propose concept containers: reusable compressed representations that preserve causal structure while minimizing compute. Concept containers function as representation-level regulation by reducing repeated recomputation, stabilizing internal trajectories, and improving transfer at equal or lower cost. We formalize containers as intervention-relevant abstractions, propose measurable criteria (causal fidelity, transfer, compute reduction), outline architectural patterns for agent systems, and describe failure modes (especially over-compression) with practical guardrails (provenance and falsifiers). Container updates are treated as committed writes to a long-lived causal cache; judges/critics should be used only to rank candidates and estimate uncertainty, and high uncertainty (tie/abstain) should defer writes in favor of evidence-gathering or falsifier generation. Includes a concrete template and an instantiated example container.