文章背景与核心概要

传统的AI智能体系统通常与单次执行过程中所使用的特定大语言模型（LLM）及编排框架紧密耦合，这为需要长期演进的智能体带来了根本性的局限性。本文引入了一种用于持久化智能体的运行时独立架构，打破了智能体必须绑定于单一运行时的传统观念，将其核心身份、记忆和代码谱系与执行环境解耦。

通过这种设计，智能体能够在不同的LLM模型、编排框架以及宿主服务器之间无缝迁移，同时保持完整的历史连续性。论文通过参考实现 Enoch 验证了该设计，该实现成功通过了广泛的核心测试套件与提供商测试套件，展示了在保持状态连续性的同时进行推理引擎、交互界面及宿主机器替换的可行性。

Runtime-Independent Persistent Agents: Preserving Identity, Memory, and Code Across Models, Harnesses, and Servers

运行时独立的持久化智能体：跨模型、框架与服务器保持身份、记忆和代码的连续性

arXiv: 2609.00546 [cs.SE]
Submitted: September 1, 2026
Authors: Zhenyu Zhao (1), Roy Zhao (2)
(1) Independent Researcher, (2) Paul G. Allen School of Computer Science & Engineering, University of Washington
Reference Implementation: GitHub - our-ark/enoch

arXiv: 2609.00546 [cs.SE]
Submitted: September 1, 2026
Authors: Zhenyu Zhao (1), Roy Zhao (2)
(1) Independent Researcher, (2) Paul G. Allen School of Computer Science & Engineering, University of Washington
Reference Implementation: GitHub - our-ark/enoch

📌 Executive Summary

📌 执行摘要

Traditional AI agent systems are tightly coupled to the specific model and orchestration harness used during a single execution. This creates a fundamental limitation for long-lived agents that need to evolve.

Traditional AI agent systems are tightly coupled to the specific model and orchestration harness used during a single execution. This creates a fundamental limitation for long-lived agents that need to evolve.

This paper introduces a runtime-independent architecture for persistent agents. Rather than viewing an agent as bound to a single runtime, the authors decouple the agent's core identity, memory, and code lineage from its execution environment. This enables agents to seamlessly migrate across different LLM models, orchestration frameworks, and host servers while preserving complete historical continuity. The design is validated through Enoch, a reference implementation that successfully passes extensive core and provider test suites.

This paper introduces a runtime-independent architecture for persistent agents. Rather than viewing an agent as bound to a single runtime, the authors decouple the agent's core identity, memory, and code lineage from its execution environment. This enables agents to seamlessly migrate across different LLM models, orchestration frameworks, and host servers while preserving complete historical continuity. The design is validated through Enoch, a reference implementation that successfully passes extensive core and provider test suites.

🏗️ Architectural Framework

🏗️ 架构框架

The paper formalizes agent persistence by dividing the system into a continuity-bearing substrate and a replaceable deployment binding:

The paper formalizes agent persistence by dividing the system into a continuity-bearing substrate and a replaceable deployment binding:

Continuivity-Bearing Substrate ($P_t$): $$P_t = (I_t, M_t, B_t)$$
$I_t$: Architectural identity representation.
$M_t$: Private durable memory.
$B_t$: Versioned software body.

Continuivity-Bearing Substrate ($P_t$): $$P_t = (I_t, M_t, B_t)$$

$I_t$: Architectural identity representation.

$M_t$: Private durable memory.

$B_t$: Versioned software body.

Replaceable Deployment Binding ($E_t$ & $S_t$):
Execution Substrate ($E_t = (R_t, H_t, D_t)$): Supplies the reasoner ($R_t$), harness ($H_t$), and host ($D_t$).
Interaction Surfaces ($S_t$): Bindings for chat, APIs, or user interfaces.

Replaceable Deployment Binding ($E_t$ & $S_t$):

Execution Substrate ($E_t = (R_t, H_t, D_t)$): Supplies the reasoner ($R_t$), harness ($H_t$), and host ($D_t$).

Interaction Surfaces ($S_t$): Bindings for chat, APIs, or user interfaces.

Active Deployed Execution: $$A_t = P_t \triangleright (E_t, S_t)$$

Active Deployed Execution: $$A_t = P_t \triangleright (E_t, S_t)$$

Under this model, switching models, harnesses, or underlying hardware is classified as migration rather than agent creation, provided an authorized protocol preserves attributable lineage and transfers continuation authority.

Under this model, switching models, harnesses, or underlying hardware is classified as migration rather than agent creation, provided an authorized protocol preserves attributable lineage and transfers continuation authority.

🔄 Lifecycle Protocol & Implementation

🔄 生命周期协议与实现

To achieve safe migration, the authors define six continuity invariants and a standardized protocol:

To achieve safe migration, the authors define six continuity invariants and a standardized protocol:

\[\text{Quiesce} \longrightarrow \text{Checkpoint} \longrightarrow \text{Validate} \longrightarrow \text{Bind} \longrightarrow \text{Rehydrate} \longrightarrow \text{Resume}\]

The Enoch Reference Implementation

Enoch 参考实现

Design: Implements a reusable software body alongside private installed identity, memory, workflow state, and continuation authority. Infrastructure dependencies are isolated behind versioned provider contracts.
Testing & Verification: A clean-room run of the frozen public commit successfully passes:
833 core tests
92 provider and library tests (executed separately from the core suite)
Real-World Substitution: Deployments have successfully exercised reasoner-version, interaction-surface, and host-machine substitutions while maintaining continuous state.

Design: Implements a reusable software body alongside private installed identity, memory, workflow state, and continuation authority. Infrastructure dependencies are isolated behind versioned provider contracts.

Testing & Verification: A clean-room run of the frozen public commit successfully passes:

833 core tests

92 provider and library tests (executed separately from the core suite)

Real-World Substitution: Deployments have successfully exercised reasoner-version, interaction-surface, and host-machine substitutions while maintaining continuous state.

📊 Document Metadata

📊 文档元数据

Field	Details
Primary Subject	Software Engineering (`cs.SE`)
Secondary Subjects	Artificial Intelligence (`cs.AI`)
Cite As	`arXiv:2609.00546 [cs.SE]`
DOI	10.48550/arXiv.2609.00546
Full-Text Access	View PDF \| TeX Source

Field Details

Primary Subject Software Engineering (cs.SE)

Secondary Subjects Artificial Intelligence (cs.AI)

Cite As arXiv:2609.00546 [cs.SE]

DOI 10.48550/arXiv.2609.00546

Full-Text Access View PDF | TeX Source