Tenure: A Local-First Proxy for Structured LLM Memory

A recent study published on arXiv (2605.11325) posits that managing memory across sessions in large language models should be viewed as a state management issue rather than a search issue. The researchers argue that similarity search is inadequate for resolving named entities in limited vocabulary scenarios, such as within a single user or an engineering team utilizing common codebases and terminology. They present Tenure, a local-first solution that upholds a typed belief store featuring epistemic status, versioned supersession, and scope isolation. By employing precision-first retrieval, Tenure integrates curated context into each LLM session, aiming to minimize re-orientation costs in workflows that are iterative and session-intensive. The paper is authored by a team of researchers and is accessible on arXiv.