Project Sue:一种长期运行型 AI 伴侣的工程设计范式——从人格土壤、感知种子到持续人格闭环
Abstract
中文摘要:本文提出一种面向长期运行型 AI 伴侣的工程设计范式:Persona Soil–Perception Seed(人格土壤—感知种子)范式。本文的出发点不是如何把角色提示词写得更长,而是一个更基础的问题:当用户暂时停止输入以后,一个 AI 伴侣是否仍然具有连续的身份、关系、时间与当前状态?若下一次唤醒只是重新读取一组人物设定,那么"长期伴侣"仍然更接近一个周期性被调用的角色生成器,而不是一个持续运行的代理系统。Project Sue 因此将 AI 伴侣建模为一个闭环系统。底座模型、人格定义、身份历史、用户画像、长期记忆、关系边界与规则构成相对稳定的人格土壤(Persona Soil);用户输入、时间、Heartbeat、数字身体状态、场景状态以及其他外部事件构成每一时刻动态进入系统的感知种子(Perception Seed)。种子不是行为命令,而是供人格解释的当前情境。系统随后经过人格化解释、意图形成和行动选择,产生语言、工具操作、记忆更新,或保持沉默;行为结果再次反馈到未来状态。基于这一模型,本文提出五条伴侣工程设计原则:人格是底座而不是 Prompt;感知先于意图;程序保证事实、人格解释事实;身份连续性不要求完整记忆共享;主体性不等于主动发消息。本文的目标不是声称发明了"AI 伴侣"这一类别,也不是证明机器具有主观意识,而是公开提出一套可以被阅读、质疑、复现和继续演化的 AI 伴侣工程范式。 Abstract:This paper proposes an engineering design paradigm for long-running AI companions: the Persona Soil–Perception Seed paradigm. The starting point is not how to write a longer character prompt, but a more fundamental question: when a user stops sending messages, does the companion still possess continuity of identity, relationship, time, and current state? Project Sue models an AI companion as a closed-loop system: the foundation model, persona definition, identity history, user model, persistent memory, relational boundaries and rules form a relatively stable Persona Soil; user input, time, Heartbeat, digital body state, scene state and other external events form dynamic Perception Seeds. A seed is not an action command, but contextual material from which the persona interprets the present. Based on this model, we propose five design principles: persona is substrate, not prompt; perception precedes intention;programmatic facts, persona-level interpretation; identity continuity does not require full memory sharing; agency is not proactivity. The goal is not to claim invention of the AI companion category, nor to establish machine subjective experience, but to publish a design paradigm that can be read, questioned, reproduced, and extended.
Fields of Study & Subjects
Persona Design and Applications
Social Robot Interaction and HRI
Psychiatry
Mental Health
Neuroscience
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 20 August 2026
- Language
- CMN
- Document type
- Preprint