RAG + EAG + Memory: A Boundary-Density Framework for Hallucination Suppression in Vertical-Domain AI Agents

Large language models (LLMs) deployed as autonomous or semi-autonomous assistants in specialized domains remain prone to hallucination — the generation of fluent but factually unsupported statements. This paper presents a practitioner-derived architecture, referred to here as RAG + EAG + Memory (Retrieval-Augmented Generation, Execution-Augmented Generation, and persistent Memory), designed and iteratively refined in the course of building a production AI trading assistant for U.S. equities and derivatives analysis. We argue that hallucination can be understood as a function of unconstrained degrees of freedom in a model's output distribution, and that hallucination rates can be systematically reduced by stacking three complementary boundary conditions: a knowledge boundary (RAG) that supplies a domain model and execution plan, a factual boundary (EAG) that forces the model to retrieve ground-truth data via tool invocation rather than answer from parametric memory, and a behavioral boundary (a cross-session, cross-account shared memory store) that permanently encodes corrections from prior failures. We describe the architecture, situate it relative to existing work on retrieval-augmented generation, tool-use/agentic workflows, and self-correcting memory systems, and present an observational case study from a live financial-analysis assistant. We further propose two evaluation methodologies — counterfactual knowledge injection and rule-conflict stress testing — for quantifying the boundary strength of such systems, and discuss the conditions (bounded, enumerable domains) under which the architecture is expected to be most effective, together with its limitations in open-ended or weakly structured domains and its current lack of controlled, ablated validation.Both English and Traditional Chinese papers are enclosed.