From Access to Derivation: Governing Behavioral Learning in Persistent AI Agents

Persistent AI agents observe human behavior within authorized data access scopes, building behavioral models that become the deepest form of platform lock-in. Current agent governance frameworks distinguish between authorized and unauthorized actions but draw no boundary between authorized data access and unauthorized behavioral derivation. An agent delegated to read email is implicitly authorized to build a predictive model of how the user prioritizes, responds, and makes decisions. No receipt is produced. No scope constrains what may be derived. No portable artifact captures what was learned.We define behavioral derivation rights as a governance primitive for persistent AI agents. Derivation rights attach to existing delegation objects, specifying whether an agent may retain behavioral patterns derived from authorized data access, for how long, and under what export conditions. The primitive extends the Agent Passport System's monotonic narrowing model: a child delegation cannot widen derivation permissions granted by its parent, and revocation of derivation rights follows the same cascade semantics as action scope revocation.We define three artifacts: (1) derivation rights and observation policy as delegation properties enforced at the gateway boundary, (2) Behavioral Memory Objects (BMOs) as signed, portable governance envelopes for behavioral attestations held by the human principal, and (3) lifecycle receipts for BMO creation, update, export, import, and deletion events. The architecture governs explicit, externalized behavioral memory artifacts, not model-internal representations or provider-side personalization pipelines. The design favors coarse, enforceable gates at architectural chokepoints over fine-grained but unverifiable inference taxonomies.We identify four open problems that the architecture does not solve: consent contagion in relational behavioral data, cognitive lock-in through bilateral human-agent adaptation, attestation fidelity measurement against strategic provider behavior, and verifiable machine unlearning. We argue that establishing this governable surface is a necessary precondition for regulatory and market enforcement, even where technical verification remains incomplete.The primitives are implemented in the Agent Passport System SDK (v1.37.0, 2,535 tests, Apache 2.0) and extend an existing IETF Internet-Draft (draft-pidlisnyi-aps-00).