What question did this study set out to answer?

This document introduces the HEEL architecture to address memory constraints in large language models (LLMs) for persistent autonomous agents.

May 11, 2026Open Access

HEEL: A State-Centric Autobiographical Memory Architecture for Persistent Agents — Beyond the Session-Centric Ontology

Key Points

This document introduces the HEEL architecture to address memory constraints in large language models (LLMs) for persistent autonomous agents.
Developed a hierarchical memory management system with four data layers: History, Experience, External, and Logic.
Established a continuous memory lifecycle of encoding, consolidation, decay, and eviction.
Decoupled agent identity into three layers: Character, Fenshen, and Yinao for fine-grained memory injection.
Demonstrated natural dissolution of engineering challenges like cross-session amnesia and personality fragmentation.
Validated memory stratification without relying on semantic labeling as per the Wall-Breaking Theorem.
Achieved a theoretical upper bound for the architecture's validity based on strategic dimensional reduction.

Abstract

Title HEEL: A State-Centric Autobiographical Memory Architecture for Persistent Agents — Beyond the Session-Centric Ontology Description This document proposes the HEEL (History, Experience, External, Logic) architecture, introducing a hierarchical virtual memory management system for LLM-driven autonomous agents. This is the first strategic dimensional ascension operation within the Agent OS theoretical framework, designed to address the first-principles constraint of finite context windows in LLMs. Core Contributions "Injectability" replaces semantics as the sole criterion for memory stratification. Data is classified into four layers—History (raw material, ℐ=0), Experience (ready-to-use finished products, ℐ=1), External (external raw material, ℐ=0), and Logic (pure strategy management layer) —based on processing state rather than semantic content. This criterion is orthogonal to data semantics, fundamentally avoiding the ambiguity and maintenance overhead of semantic classification. Establishes a lifelong memory repository ℳ with a complete metabolic closed loop. By Axiom 0, the Agent possesses a persistent memory repository whose existence does not depend on any specific Session. By Axiom 2, this repository undergoes continuous Encoding → Consolidation → Decay → Eviction, preventing infinite growth from degrading retrieval efficiency. The architecture achieves "perpetual context abundance" by injecting only highly-ready Experience-layer data into the context window through the Concentration Function fconcentration. Three-layer identity decoupling: Character-Fenshen-Yinao. The Agent's identity is decoupled into three orthogonal layers: Character (persistent personality bound to ℳ), Fenshen (temporary view under a specific task, a concrete instantiation of fconcentration), and Yinao (hot-swappable LLM driving inference). This transforms memory injection from a coarse-grained inheritance model (Session inherits all or nothing) to a fine-grained composition model (Fenshen composes relevant Experience subset on demand). Paradigm dissolution, not problem-solving. We demonstrate that the "engineering challenges" under the old Session-Centric Ontology—cross-session amnesia, personality fragmentation, context management burden—are not solved but naturally dissolved: they cease to be raised because the architectural premises that generated them (Session as memory boundary, semantic classification) have been removed. Theoretical Status Upgrade (Updated May 2026) Since the initial release of this document, the Agent OS theoretical framework has achieved a critical formal closure. The Wall-Breaking Theorem, which provides a rigorous proof of the cost of strategic dimensional reduction, now offers an unbreakable theoretical upper bound and an unavoidable mathematical guarantee for this architecture: The engineering intuition of this architecture (replacing semantic labels with processing state as the classification criterion) constitutes the engineering starting point for the diagnosis of strategic dimensional reduction formalized in the Wall-Breaking Theorem. Conversely, the rigorous formal proof of the Wall-Breaking Theorem provides this document's architectural choice—the separation of readiness inspection from triggerability management—with a precise theoretical upper bound and a logical necessity argument. The two are mutually reinforcing: this document serves as the engineering intuition source and constructive case for the theorem; the theorem serves as the theoretical foundation and necessity guarantee for this architecture. Citation Note The formal guarantee of this architecture's engineering validity derives from the Wall-Breaking Theorem's rigorous proof of the cost of strategic dimensional reduction. This architecture is cited as a constructive case in the Wall-Breaking Theorem to provide intuitive anchoring and historical context, and does not constitute a logical premise of that theorem. HEEL is the first engineering-dimensional unfolding of the Agent OS theoretical system. For the complete worldview of this system, see the Agent OS Paradigm Manifesto. Version v1. 1. 0 Update Log v1. 1. 0 (May 2026): Updated terminology from "attention" to "concentration" to distinguish from Transformer's attention mechanism. Added PDF version to ensure LaTeX formula rendering. Corrected formula display issues. 中文摘要本文档提出HEEL (History, Experience, External, Logic) 架构, 为LLM驱动的自主Agent引入分层虚拟内存管理系统。这是Agent OS理论体系中的第一次策略升维操作, 以"可注入性"替代"语义"作为记忆分层的唯一判据, 建立终身记忆库ℳ及其代谢闭环, 并通过"角色-分身-义脑"三层身份解耦, 实现从粗粒度继承到细粒度组合的记忆注入范式转变。旧范式下的"工程难题"——跨会话失忆、人格撕裂、上下文管理负担——在新公理体系下被自然消解, 而非被解决。 2026年5月更新: 《破壁定理》的严格形式化证明, 为本架构的"就绪性审查独立于触发度管理"提供了不可突破的理论上限。本文是定理的工程直觉来源与构造性案例;定理是本文的理论根基与必然性保证。

HEEL: A State-Centric Autobiographical Memory Architecture for Persistent Agents — Beyond the Session-Centric Ontology

Key Points

Abstract

Cite This Study