The Agents Epistemic Baggage

Agentic AI systems with writable persistent memory do not only retrieve information. They also store their own summaries, conclusions, plans, and execution histories. When a system loses the distinction between received information and its own inferences, an inference may acquire or retain a status that its evidence does not justify. Such a record may then influence decisions or actions, even through policy-permitted use. Unauthorized authority escalation is a separate control failure that the project also examines.

This project develops a risk model for this drift process, called self-reinforcing epistemic drift, and proposes how to make it auditable. It describes two drift mechanisms: a re-storage loop, where agent-generated material is stored, retrieved, used, and written back, and consolidation drift, where summarization, abstraction, or compression strips provenance, validation state, temporal validity or evidential support from a record that then replaces the original or gains priority over it.

The project provides:

• a logical audit schema that keeps representation type, provenance, validation state, temporal validity, evidence relation, change history, epistemic transition basis, authority basis, and usage and decision lineage separate;

• an operational authority profile that distinguishes permitted authority, observed exposure, recorded use, and causal influence;

• a lifecycle test protocol, including tests for whether justified updates and re-review remain possible;

• an audit operationalization with typed audit objects, transition rules, controls, provisional metrics, and reproducible test specifications;

• a user guide for running an audit.

The model, schema, and controls are proposed constructs for testing, not established mitigations. Contributions of empirical research, critiques, schemas, controls, test cases, and framework mappings are welcome.

License: CC BY-SA 4.0 https://creativecommons.org/licenses/by-sa/4.0/