AI Loses Codebase Context Between Tasks — Researcher Proposes Semantic Memory Layer

A developer and researcher has highlighted a core limitation of AI coding agents: the understanding they build while navigating large codebases is lost once a session ends, forcing repeated investigation in future tasks. To illustrate the problem, the author uses the example of moving email delivery to an async queue, where a single function serves both two-factor authentication and newsletters, each with distinct timing and consent requirements. Without retained context, an AI agent must re-spend significant tokens reconstructing dependencies and design assumptions from scratch each time. The researcher proposes building a persistent semantic layer — grounded in a mathematical framework called Algebraic Architecture Theory — that would store concepts, relationships, and contextual rules in a reusable, shareable form. The goal is to let both humans and AI agents query this layer to assess how proposed changes affect system behavior without re-reading the entire codebase.