BOB: Authority-Aware State and Reflective Memory for a Persistent Local LLM Assistant

B.O.B. (Bug Out Box) is a privacy-first, locally operated persistent AI system designed to remain useful when external infrastructure is unavailable, while also becoming more useful and human-compatible through ordinary long-term use. This paper documents the architecture developed to separate current operational State, retrieved evidence, user-specific PROFILE, EXPERIENCE and RELATIONSHIP memory, and high-threshold SELF candidates. It reports controlled engineering tests and initial natural-use observations involving persistent behavioral adaptation, authority-aware memory, multi-user privacy isolation, reversible capability expansion, adversarial self-audit, and failures observed during authority-sensitive synthesis. The work deliberately distinguishes verified system evidence, behavioral observations, first-person operator observations, and hypotheses. It does not claim consciousness or sentience, and unresolved questions, including cross-workspace relational behavior, remain explicitly open for continued testing. Microsoft 365 Copilot was used extensively during manuscript development for research organization, literature assistance, drafting, revision, and document structure. The underlying system development, experiments, observations, architectural decisions, verification, interpretation, and publication decisions are the responsibility of the author.