Papers
Research toward an AI that can live among people for years, learn from that life rather than from a lab, and grow more capable without growing beyond what it was trusted with. Memory and authority are where the work begins.
-
Authorization Before Context: A Model-Neutral Audience Boundary Against Cross-Audience Memory Leakage in Agentic Systems
An AI assistant learns something from one person and can later repeat it to another. Relevance-only retrieval, output filters and separate memory silos all fail here. The paper proposes one rule: a memory reaches the model only if everyone in the current conversation was there when it was learned. When the audience is unclear, nothing private gets in. Four properties are proved. Across 79 synthetic scenarios, zero forbidden inclusions.
The evidence is synthetic, and the paper says so. It makes no production claims and is not a general defense against prompt injection.
-
A second paper, preprint forthcoming
Details here once the preprint is posted.