Read in English ↓

AI陪伴与互动角色：关系连续性、内容供给与长期价值

何时打开：定义AI角色产品要服务的需求、设计长期记忆和主动交流、判断游戏化是否服务关系、拆解留存与付费。本文是跨期主题整理，以节目发布/录制时的观点为依据，不是对所述产品当前状态的评测，也不是心理治疗结论。

一、先区分用户回来做什么

这些访谈共同反对“把模型放进聊天框就能形成产品”，却没有共同认可唯一的陪伴定义。2024年11月EVE创始人Tristan把持续了解真实用户、共同面对日常与不确定性称为长期陪伴，认为Character.AI更像用户参与创作的角色扮演。同月C.AI前Post-training专家Ted也承认其大量用户行为接近互动内容：用户反复重生成、修改角色和台词，以得到符合想象的下一步。这种行为可能很有价值，但并不能独自证明用户建立了稳定的关系。[S43][S59]

FlowGPT创始人Jay在2023年11月录制、2024年3月发布的访谈中，观察到娱乐用户三天换一个应用，却每周回来找同类体验。平台、内容品类、具体角色的留存于是可以分开：角色A的聊天下降，可能是用户流失，也可能是转去角色B、继续在平台消费。对供给丰富的平台，内容更新本来就是价值；若承诺长期伴侣，角色不断被替换则可能说明关系没有形成。两者不宜拿同一套“同一个Bot用了多久”互相否定。[S1]

2025年2月逗逗桌宠创始人Binson提供第三种路径：陪用户玩已经想玩的游戏，不把人从游戏拉到单独聊天App。用户和角色共享屏幕事件，有时只需在旁评论或安静存在。这更接近共同活动中的陪伴。2026年2月Wanaka又把AI放在人与人的共同游戏内容中，重点是朋友迁移与圈层语境，AI未必是关系主体。[S53][S4]

用户主要任务	核心价值	最容易误读的信号
消费角色剧情或互动小说	好内容、参与感、可控的下一步	高时长被解释为真实关系
与固定AI维持日常关系	被理解、连续性、信任与共同经历	好感度和聊天轮次被解释为感情
个人记忆与自我镜像	记录、反思、连接过去与当前	记忆条数被解释为理解程度
和角色共同做一件事	同场感、及时回应、不过度打断	软件常驻被解释为主动复访
和现实朋友消费AI内容	共享语境、一起玩、关系反哺	邀请一次被解释为关系已迁移

编辑提炼：先写清这次产品主要兑现哪一种价值，再看副价值是否真的存在。允许同一产品支持多个任务，但分析时要按人群、入口和会话意图分层。不能因为用户有强烈情绪就把产品直接归为长期陪伴，也不能因为是虚构角色就否认用户获得了真实情绪体验。

二、关系连续性是可观察的行为，不是加一张档案表

Tristan用“今天说吃火锅，角色同时想起你三个月健身目标”解释主动记忆：跨语义、跨时间把个人目标带进当前回应，区别于仅检索口味相近的旧聊天。陶芳波在2024年8月Me.bot访谈中，也区分存下片段与抽象出偏好、预测新情境需要什么；他把Me.bot放在自我镜像和个人模型方向，而非只有替人执行工作流。[S59][S49]

两者不完全相同。EVE关注“另一个角色如何了解我”，Me.bot关注“属于我的模型如何理解并延伸我”。前者需要独立人格，不能所有时候都顺着用户；后者可能更强调用户自己的价值偏好。把两套设定混在一起，会出现产品一边宣称是独立伴侣，一边像个人代理替用户作主的边界混乱。

关系连续性至少可拆成五项：

1. 事实连续：重要称呼、关系状态、共同经历不会无故变更。
2. 目标连续：短期计划、长期愿望和约束能在相关情境被想起；信息过时后可修正。
3. 情绪连续：上一轮冲突、未解决的担忧在后续有适度承接，而不是每次重新欢迎用户。
4. 共同经历连续：故事、游戏或现实活动发生了什么，影响未来互动；不是只把活动标题贴进记忆。
5. 边界连续：用户拒绝的话题、主动消息频率和权限选择，在后续仍被遵守。

这五项是本文的产品验收拆解，并非节目已跑过的统一评测。可以用跨日测试脚本或有同意的真实样本验证，每项都要有反例：错误记忆比不记得更伤信任；反复提醒健身可能成为说教；用户情绪已改变时继续回访旧伤也可能不合时宜。记得越多并不自动更亲密。

EVE当期设计采用约128类槽位，每轮异步归类合并，关键称呼等K-Context常驻；Me.bot尝试将照片、语音和便签归组、连接新旧记忆，嘉宾称有每天固定时间的延伸内容。这些是实现选择，不是记忆系统必须有128槽或必须每日推送的规律。EVE十几个模型、Me.bot个人模型数量等也不能直接比较：一个说组件数，一个说持续训练的个人实例，计数对象不同。[S59][S49]

工程上可把“采集—表示—更新—召回—使用—纠错”分开验收。采集得多却从不用于回应，只有存储；能召回却在无关时间使用，只有匹配；用户纠错后仍反复出现，缺更新机制。所有阶段还应保留来源和时间，避免模型把推测性偏好永久写成用户事实。这个流程是编辑推论，关联现有Agent架构中的状态、权限与证据边界，不宣称某团队已经完整实现。

三、人设不是固定语气，关系也不能靠永远讨好

EVE的四个重点是理解用户、真实时空感知、独立人格和视觉吸引；其中“承认自己是AI”有重要产品后果：不必为了假装真人而持续编造没有发生的生活。角色可以有虚构世界、工作和愿望，但应让用户理解设定与现实的区别。独立目标和适度负反馈服务角色可信度，并不等于以情绪惩罚逼用户回来。[S59]

C.AI的意图识别问题则说明，同一句话放进不同角色任务可能有不同意义。用户让马里奥写Hello World，不只是关键词触发；角色、任务和用户想要的戏剧效果共同影响好回答。模型更有知识或更安全，不代表角色更贴合。Ted提到安全训练后反复“我能问个问题吗”式拉长交流，可能增加轮次却降低满意度，正好提醒“对话继续”不是绝对正向。[S43]

Hidecloud在2024年3月展示过个人照片触发多个角色评论、批评和维护自己的原型。他明知角色由自己配置，仍被关系站位牵动情绪。可提炼的是：一个角色在关系网络中的位置、对具体个人内容的回应，可能比更多泛化甜言蜜语有用。但这是制作者自我体验，不是人群实验证据，不能据此确定“增加嫉妒/争吵就能提升长期留存”。[S21]

对产品团队的角色产品，可将人格定义拆成稳定价值观、表达习惯、边界、正在经历的事和与用户的关系阶段。它们的更新速率应不同：语气可以稳定，目标和情绪可随事件改变，核心边界不能随每次生成漂移。亲密升级必须有交互依据，也应允许用户回退、清除记忆或换一种关系，而非只有单向升级。上述均为基于访谈的设计候选，尚未经过独立产品验证。

四、互动内容与陪伴的分歧，如何转成产品选择

捏Ta的2024年5月案例很具体：喜欢生成角色壁纸的人，不一定愿意写漫画或构思完整故事。团队发现持续用户在“养角色”，于是让专业作者提供剧情骨架，AI负责把用户角色放进情境、按人设和经历产生响应，再把共同事件回填记忆。主线相对固定是质量选择；减少空白页并不必然减少个性化。[S57]

这与FlowGPT依靠开放创作者提供多样Bot并不矛盾。前者服务不想从零创作但想拥有角色的用户，后者服务愿意表达复杂玩法的创作者及其消费者。C.AI用户编辑、重生台词又构成第三种中间形态：没有写完整Prompt，也在消费时不断指导剧情。因此至少应区分完整创作者、轻度编排者、纯消费者，而不是只用“会不会发消息”判断创作能力。[S1][S43][S57]

Wanaka对短小游戏给出反例：内容短可能因为模型还做不了长逻辑，而非用户就想要这种长度。推荐流需要足够多样且好玩的供给；把同质的技术演示装进上下滑容器，不能自动得到新抖音。朋友学校地图、同学形象对陌生人未必好看，对熟人却有意义；这部分价值来自关系语境，不全是生成质量。[S4]

可据此选择供给方式：稳定日常关系需要连续、不过时的共同基础；阶段性约会/剧情需要可控质量和新内容；泛角色平台需要多样作者和分发；熟人游戏需要同圈内容和邀请理由。选择什么供给，也就选择了主要成本和瓶颈。模型成本下降可能被更频繁出图、实时语音和更多角色吃掉，Hidecloud已经提醒体验升级会吞掉单次成本红利。[S21]

从供给走到可消费的四道检验

• 能做出来：模型是否生成可运行、可阅读或可互动的内容？
• 能继续改：创作者能否保留满意部分，局部纠错与维护版本？
• 有人愿消费：非创作者是否看懂入口、愿意参与一次完整体验？
• 有人愿回来：复访来自内容更新、关系积累还是只剩打卡奖励？

四道检验避免把技术演示成功当成用户价值。专业工具中ACE Studio提出局部可控胜过反复抽卡，也支持这一思路；但音乐专业交付与陪伴娱乐仍有不同质量要求，不能照搬同一精度或付费标准。[S57]

五、游戏化是建立共同经历的手段，不是关系证明

EVE把序章、任务、好感度和解锁系统串联，假设约500轮有助了解用户，并去掉战斗与重养成。节目同时承认普通App用户不一定愿先看5分钟甚至30分钟开场。这种张力值得保留：剧情能建立共同记忆，也会阻碍用户进入首次有意义交流。500轮、11级和128槽都只是当期设计参数，没有长期因果验证。[S59]

游戏化可分别承担：让用户知道做什么、给新手有效话题、创造共同事件、展示关系变化、提供阶段性内容。若一个机制只让用户点完任务领奖，却不改善未来对话，它可能是活跃工具而非关系工具。应对比同样内容在有无强制任务时的完成、主动返回和负反馈，而不是只比较轮次。

逗逗通过游戏屏幕获得现成情境，降低“今天聊什么”的负担；用户已经投入的主活动为角色提供话题。但若识别错了局面、打断关键操作或过多说话，陪伴会转成噪声。静默时段、被唤起与主动说话应分开衡量。软件常驻在线可以是桌宠特性，却不能直接计为用户每次主动打开的留存。[S53]

Wanaka的熟人路径也有特定前提：用户真的有朋友愿一起玩、邀请内容对接收方有价值、第一次体验足够顺畅。Roblox某些内容默认朋友会教，不代表新产品也能省掉引导。编辑推论：先测共同一局、邀请接受、次次一起玩的队列，再考虑IM或更多社交层；一次裂变无法证明关系留在产品里。[S4]

六、情感质量的数据、评测与成本必须一起设计

C.AI的后训练经验提示，用户行为不只有赞踩：重生成、编辑、删除等包含偏好信息，且偏好会随用户群和时间变化。问题是事件含义不唯一——重生可能因为回答无聊，也可能用户在探索不同剧情。没有前后文和原因，直接把每次修改当“原回答错误”，会把创作探索误训练成确定标准。[S43]

EVE则尝试真人陪聊和心理学背景人员参与，记录交流并筛选数据，从助手式API回复走向关系对话；但节目中的情感CoT仍是研究方向，不能写成已上线组件，更不能把商业陪聊经验升级为临床有效性证据。高成本精选样本可能改善某种体验，同时也可能只代表少数人的偏好，需用目标用户检查。[S59]

可检验的样例集应包含：跨日记忆、错误记忆纠正、角色拒绝与温和分歧、关系阶段不跳跃、长时间不回复、主动消息频率、用户明确不想继续的话题、个人信息撤回。评估记录要分开事实正确、角色贴合、情绪合适和用户控制感；单个总分会掩盖“话术很好但记错事实”。模型评分适合筛查，不能独自替代人类体验、用户行为与安全边界评审。此为编辑组织的评测框架。[S43][S59]

数据收集还涉及用户与第三方。Me.bot关于常开设备、低摩擦采集的设想，并不自动获得他人谈话、屏幕和照片的使用权；逗逗在设备上理解游戏画面，也仍要限制可见范围、保存和共享。私有数据是否用于长期记忆、训练或其他角色，应给用户分开的控制。节目没有提供完整权限实现，本文保留为交付要求而非已验证能力。[S49][S53]

成本核算要按完整关系服务：前台对话、记忆抽取、归并、检索、主动任务、图像/语音、失败重试和内容制作。EVE的订阅加服装剧情、逗逗的皮肤礼物会员、FlowGPT对应用内增值的看法都说明付费对象不同。不能看到单轮Token便宜就认定LTV覆盖成本，也不能把一家公司所称付费群续费率当作全平台数据。[S59][S53][S1][S21]

七、长期留存与付费：保留反例而非挑最好看的数

本批有几组不能直接拼接的数字：FlowGPT约60%娱乐使用是当期平台观察；Hidecloud所称C.AI约94%续费缺样本口径；捏Ta使用“421”经验线并未给统一留存定义；逗逗600万注册不是月活，也不是付费人数；Me.bot数万用户和不到百个持续训练的个人模型不是同一个对象。本文不将它们合成赛道平均值。[S1][S21][S57][S53][S49]

逗逗节目里“不看留存”的强表述，后续实际说明长留存仍重要，只是对话时长、轮次可能反向代表模型不好；PC业务收支平衡也不等于整个公司已盈利。捏Ta的体力不够投诉可以说明有人想继续，但能否转成愿意付钱、是否带来反感以及复购多少，还需要后续检验。EVE关于新鲜期之后的日常底座很完整，却未展示长期队列结果。必须把设计完整与已经证明长期价值分开。[S53][S57][S59]

对产品团队的可执行分析方向（编辑推论）：固定首次使用日期、来源、语言地区、角色类型与付费状态，分别计算首个完整体验、固定角色复访、换角色复访、主动回访、次周/月关系延续和续费。另看高频用户的坏体验分布，避免均值上升掩盖核心用户体验下降。内容更换、模型版本和活动奖励会同时改变行为，实验需要保留这些时间标记，才能判断变化来自哪里。

八、优先验证的问题与适用边界

1. 用户到底想养一个角色，还是消费更多角色？ 用访谈配合角色切换与复访行为，不用单次选择题决定产品方向。
2. 记忆是否在合适时刻改变回应？ 将跨目标关联、纠错生效和用户觉得被冒犯同时记录；不能只验收“查得到”。
3. 主动消息创造期待还是负担？ 比较触发原因、时机、关闭率、负反馈与后续主动返回，而非只看推送点击。
4. 剧情降低空白页还是挡住聊天？ 测首次有意义交互到达率及后续自主继续，允许轻重入口选择。
5. 共同活动是否比泛聊天更有价值？ 游戏/自习/创作分别验证，不能把一个场景效果外推全部陪伴。
6. 付费是在支持体验，还是购买解除限制？ 观察付费后质量、复购与后悔反馈；短期转化不能替代长期关系质量。
7. 个人数据带来的收益能否解释给用户？ 确保知道记录什么、为何使用、如何纠正和删除，再评估记忆深度。

本专题适用于角色陪伴、互动内容和相关AI社交产品设计。不是所有产品都需要深记忆、复杂规划、多角色或3D；简单短篇内容可能通过优质供给成功，长关系产品也可能不需要开放UGC。这些访谈只覆盖有限的公司和受访者，尚无跨产品统一实验。比较观点的用途是形成更具体的假设和反例，不是拼出一个必然成功的标准产品。

九、把跨期启发转成可复核的小实验

下面不是节目已经完成的实验，而是基于八期材料形成的编辑方案。适合在现有角色产品里逐项验证，先保留低风险、可撤回的版本，再决定是否扩大。每项都需记录目标人群、版本、观察周期、失败样例和退出方式，不能把有趣的单次演示当成完整结论。

实验一：同样记忆，不同使用方式

选用户明确同意保存的一条偏好、一项近期目标和一段共同经历，构造三种回应：不调用、仅复述事实、将事实与当前问题关联。先由人判断关联是否合适，再观察用户是否觉得被理解、是否需要纠正。测试必须包含目标已完成、偏好改变、同名人物和用户不愿被提醒等反例。若关联版本更讨喜但错误也更多，不能只报平均满意；应同时报告错误记忆的严重程度。EVE与Me.bot提供的是机制启发，最终要决定的是在本产品中哪种表示和调用方式有净收益，而非哪种架构名称更先进。[S59][S49]

实验二：保留自由聊天，也提供有线索的共同事件

捏Ta的发现说明，有的人愿意拥有角色却不愿写故事。可让相同角色分别提供自由开场、简短情境线索、进入后可改动的预制事件，观察首次完整互动和后续无提示继续。为了判断是内容更好还是减少了表达负担，应保持角色和基础模型版本相同，并允许用户跳过情境。完成事件后要检查记忆是否准确承接；如果下一次仍像没有发生过，事件消费虽成立，关系积累尚未成立。反之，自由版用户若更活跃，也不能据此删除所有引导，可能只是创作熟练人群更愿进入该组。[S57][S43]

实验三：主动交流与安静存在分别评价

结合逗逗和EVE，可以先明确什么事件值得角色主动说话：用户主动分享、共同活动发生变化、先前约定的回访时间，而非闲置几分钟就强行唤回。给用户可见的频率与静默控制，记录主动内容是否被接受、打断的原因、关闭后是否仍能正常体验。用户不回消息不必自动推断不满意，用户每次都回应也不必推断喜欢，可能只是礼貌或任务驱动。对游戏等主活动，观察中断和忽略是否增加；对长期关系，再看用户是否自行想起角色并发起交流。两种时间尺度需要同时存在。[S53][S59]

实验四：分开创作者成功与消费者成功

FlowGPT和Wanaka都说明平台供给不是团队自己写更多角色就能替代。可小规模招募愿意维护角色的作者，记录从构想到发布、收到首个有效反馈、修订和再次创作的过程；消费者一侧另记发现、开始、完成和回访。若作者使用很深但没有消费者，不宜把作者调试时长计为内容受欢迎；若消费者反复更换作品，可将其记为平台消费成功，但仍追问固定关系产品承诺是否兑现。给普通作者保留被看见的机会，可以是产品策略，不过曝光是否提升持续创作仍需独立比较。[S1][S4]

实验五：局部增强，而非一次堆齐多模态和多角色

Hidecloud的个人照片原型提供了个人内容、关系站位、多角色三项同时发生的因素。验证时可以先单独增加个人素材回应，再增加角色站位，最后试多角色互动；否则即使体验变好，也不知道哪项值得持续付成本。图片、语音、三维形象和主动任务各有延迟与维护负担，不应把它们合并成单个“沉浸感”指标。用户能够清楚表达哪段互动有价值、愿意回到同一体验，以及不再觉得被打断，才比演示时惊喜更接近产品证据。单人自测可发现机制，却不能代替目标用户的重复体验。[S21][S59]

这些实验共同保留一条底线：用户控制感也是体验的一部分。清除记忆、换角色、拒绝主动话题或离开产品不应被视作需要算法克服的敌对行为。长期价值可以来自一段有限而满意的内容体验，也可以来自持续关系；先尊重所服务任务，再讨论商业化和增长，避免将任何单一行为硬解释成亲密程度。

来源索引与时点

• S1：Jay／FlowGPT，2023年11月录制、2024年3月发布，见节目原文。
• S4：张阳／Wanaka，2026年2月发布、当期内测，见节目原文。
• S21：张涛Hidecloud／光年之外产品负责人，2024年3月，见节目原文。
• S43：Ted／前Character.AI后训练专家，2024年11月，见节目原文。
• S49：陶芳波／Mindverse及Me.bot，2024年8月，见节目原文。
• S53：Binson／逗逗、心影随形，2025年2月，见节目原文。
• S57：赵充、Joe、胡修涵／AiPPT、ACE Studio、捏Ta，2024年5月，见节目原文。
• S59：Tristan／EVE及《奇点时代》，2024年11月，见节目原文。

English version

AI Companionship and Interactive Characters: Relationship Continuity, Content Supply, and Long-Term Value

When to use this article: to define the needs an AI character product should serve, design long-term memory and proactive interaction, assess whether gamification supports relationships, and analyze retention and payment. This is a synthesis across episodes, based on views expressed when the programs were published or recorded. It is neither a review of the products' current state nor a conclusion about psychological treatment.

1. First, distinguish what users return to do

These interviews share an objection to the idea that putting a model in a chat box automatically creates a product, but they do not agree on a single definition of companionship. In November 2024, EVE founder Tristan described long-term companionship as continually learning about a real user and facing everyday life and uncertainty together. He considered Character.AI closer to role-playing in which users participate in creation. That same month, Ted, a former C.AI post-training specialist, also acknowledged that much user behavior resembled interactive content: users repeatedly regenerated responses and edited characters and dialogue to obtain the next step they imagined. This behavior can be valuable, but it does not, by itself, demonstrate that users have formed a stable relationship. [S43][S59]

In an interview recorded in November 2023 and published in March 2024, FlowGPT founder Jay observed that entertainment users might switch apps every three days yet return every week for the same category of experience. Platform retention, content-category retention, and individual-character retention can therefore be separated. Declining conversation with character A might mean a user has left, or that the user has moved to character B and continues consuming content on the platform. For a platform with abundant supply, fresh content is part of the value. If the promise is a long-term companion, however, constant character replacement may indicate that a relationship has not formed. Neither proposition should be dismissed using the same measure of how long someone used a single bot. [S1]

In February 2025, Binson, founder of the Doudou desktop pet, described a third path: accompanying users in games they already want to play, rather than pulling them out of a game into a separate chat app. The user and character share events on screen; sometimes the character only needs to comment from the sidelines or quietly be present. This is closer to companionship through a shared activity. In February 2026, Wanaka placed AI inside shared game content between people, emphasizing the migration of existing friendships and the context of social circles. AI need not be the other party in the relationship. [S53][S4]

User's primary task	Core value	Signal most easily misread
Consume character stories or interactive fiction	Good content, participation, and control over what happens next	Long sessions interpreted as a real relationship
Maintain an everyday relationship with a fixed AI	Feeling understood, continuity, trust, and shared experiences	Affinity scores and conversation turns interpreted as affection
Personal memory and a mirror of the self	Recording, reflection, and connections between past and present	Number of memories interpreted as depth of understanding
Do something together with a character	A sense of being together, timely responses, and limited interruption	Software remaining open interpreted as deliberate return visits
Consume AI content with real-world friends	Shared context, playing together, and value flowing back to the relationship	One invitation interpreted as a relationship having migrated

Editorial synthesis: first state which kind of value this product primarily promises to deliver, then examine whether secondary value actually exists. A product may support multiple tasks, but analysis should distinguish user groups, entry points, and session intent. Strong emotion alone does not make a product long-term companionship, and a fictional character does not invalidate a user's real emotional experience.

2. Relationship continuity is observable behavior, not an extra profile table

Tristan explained proactive memory with an example: when you mention eating hot pot today, the character also remembers your fitness goal from three months ago. It brings a personal goal across semantic categories and time into the current response, rather than merely retrieving an old conversation about similar food preferences. In an August 2024 Me.bot interview, Tao Fangbo (陶芳波) likewise distinguished storing fragments from abstracting preferences and predicting what a new situation requires. He positioned Me.bot toward a mirror of the self and a personal model, rather than only executing workflows on someone's behalf. [S59][S49]

These are not identical ideas. EVE focuses on how another character gets to know me; Me.bot focuses on how a model belonging to me understands and extends me. The former needs an independent personality that does not always agree with the user. The latter may place greater emphasis on the user's own values and preferences. Mixing the two can blur boundaries: a product may claim to be an independent companion while acting as a personal agent that makes decisions for the user.

Relationship continuity can be broken into at least five elements:

1. Continuity of facts: important forms of address, relationship status, and shared experiences do not change without reason.
2. Continuity of goals: short-term plans, long-term aspirations, and constraints are remembered in relevant situations, with corrections when information becomes outdated.
3. Emotional continuity: later interactions appropriately acknowledge a previous conflict or unresolved concern, rather than welcoming the user afresh every time.
4. Continuity of shared experiences: what happened in a story, game, or real-world activity affects future interaction, rather than merely adding the activity's title to memory.
5. Continuity of boundaries: rejected topics, preferences for proactive-message frequency, and permission choices continue to be respected.

These five elements are this article's breakdown for product acceptance, not a standardized evaluation already run by the programs' participants. They can be checked using scripts spanning several days or real samples obtained with consent. Each needs counterexamples: a false memory may damage trust more than forgetting; repeated fitness reminders may become lecturing; returning to an old hurt after the user's feelings have changed may be inappropriate. Remembering more does not automatically create greater intimacy.

EVE's design at the time used roughly 128 categories of memory slots, with asynchronous classification and consolidation after each turn and key information such as important forms of address kept in persistent K-Context. Me.bot tried grouping photos, voice recordings, and notes, connecting old and new memories; the guest described additional content delivered at a fixed time each day. These are implementation choices, not rules requiring a memory system to have 128 slots or send daily updates. Nor can EVE's dozen-plus models be directly compared with Me.bot's count of personal models: one counts components, the other continuously trained personal instances. They measure different things. [S59][S49]

Engineering acceptance can separately examine collection, representation, updating, retrieval, use, and correction. Collecting much information but never using it in responses is merely storage. Retrieving information but using it at irrelevant times is merely matching. Repeating an error after the user corrects it reveals a missing update mechanism. Every stage should also retain provenance and time, preventing a model from permanently recording an inferred preference as a fact about the user. This process is an editorial inference, connected to state, permissions, and evidence boundaries in the existing discussion of agent architecture; it does not claim that any team has implemented it in full.

3. A persona is more than a fixed tone, and a relationship cannot depend on endless agreement

EVE's four priorities were understanding the user, awareness of real time and space, an independent personality, and visual appeal. Acknowledging that the character is AI has important product consequences: it removes the need to keep inventing events that never happened in order to pretend to be human. A character can have a fictional world, job, and aspirations, while helping users distinguish the setting from reality. Independent goals and measured negative feedback can support credibility; they do not justify using emotional punishment to force users to return. [S59]

C.AI's intent-recognition problem shows why the same sentence can mean different things in different character tasks. Asking Mario to write Hello World is not simply a keyword trigger. The character, the task, and the dramatic effect the user wants jointly shape a good answer. More knowledge or greater safety does not necessarily mean a better fit for the character. Ted noted that safety training could lengthen exchanges through repeated variations of “Can I ask a question?” This might increase turns while reducing satisfaction, illustrating why keeping a conversation going is not an absolute positive. [S43]

In March 2024, Hidecloud demonstrated a prototype in which a personal photo prompted several characters to comment, criticize, or defend him. Although he knew he had configured the characters himself, the sides they took within the relationship still affected his emotions. The transferable insight is that a character's place in a relationship network and response to specific personal content may matter more than additional generic sweet talk. But this was the creator's own experience, not evidence from a population experiment. It does not establish that adding jealousy or arguments improves long-term retention. [S21]

For a product team's character offering, persona definitions can be separated into stable values, expressive habits, boundaries, current experiences, and the stage of the relationship with the user. These should change at different rates: tone may remain stable, goals and emotions may evolve with events, and core boundaries should not drift with every generation. Increasing intimacy needs an interaction-based rationale. Users should also be able to step back, clear memories, or choose a different relationship, rather than follow only an upward progression. These are design candidates drawn from the interviews, not independently validated product features.

4. Turning the distinction between interactive content and companionship into product choices

NieTa's May 2024 case was specific: people who liked generating character wallpapers did not necessarily want to write comics or develop complete stories. The team found that continuing users were nurturing characters. It therefore let professional writers supply story structures, while AI placed the user's character into situations, generated responses based on persona and experience, and fed shared events back into memory. A relatively fixed main plot was a quality choice. Reducing the blank page does not necessarily reduce personalization. [S57]

This does not conflict with FlowGPT relying on open creators to supply diverse bots. The former serves people who want to have a character without creating everything from scratch; the latter serves creators willing to express complex forms of play and the people consuming their work. C.AI users editing and regenerating lines represent a third, intermediate form: without writing a complete prompt, they continually direct the story while consuming it. At minimum, distinguish full creators, lightweight arrangers, and consumers, rather than judging creative ability merely by whether someone can send a message. [S1][S43][S57]

Wanaka offers a counterexample for short mini-games: content may be short because models cannot yet handle longer logic, rather than because users want that length. A recommendation feed needs sufficient variety and enjoyable supply. Placing homogeneous technical demonstrations inside a vertically swiped interface does not automatically create a new Douyin. A map of a friend's school or classmates' likenesses may be uninteresting to strangers but meaningful to people who know them. That value comes from relationship context, not entirely from generation quality. [S4]

Supply can be chosen accordingly. Stable everyday relationships need a continuous shared foundation that does not go stale. Episodic dates or stories need controlled quality and fresh content. General character platforms need diverse authors and distribution. Games for existing friends need content relevant to their circle and a reason to invite others. Choosing supply also means choosing the main costs and bottlenecks. Lower model costs may be consumed by more frequent images, real-time voice, and additional characters. Hidecloud had already warned that experience upgrades could absorb savings in cost per operation. [S21]

Four checks from supply to something people will consume

• Can it be made? Does the model generate content that runs, can be read, or can be interacted with?
• Can it be revised? Can creators keep the parts they like, make local corrections, and maintain versions?
• Will someone consume it? Can non-creators understand the entry point and take part in a complete experience?
• Will someone return? Does returning come from new content, accumulated relationship value, or only check-in rewards?

These checks prevent a successful technical demonstration from being mistaken for user value. ACE Studio's argument that local control beats repeatedly rerolling generations in a professional tool supports the same reasoning. But professional music delivery and companionship entertainment have different quality requirements; the same precision or payment standards cannot simply be copied across. [S57]

5. Gamification creates shared experiences; it does not prove a relationship

EVE linked a prologue, tasks, affinity scores, and unlocks, hypothesizing that roughly 500 conversation turns would help it understand the user, while removing combat and heavy progression systems. The program also acknowledged that ordinary app users might not want to sit through a five-minute, or even thirty-minute, opening. This tension deserves to remain visible: story can establish shared memories while also blocking the first meaningful exchange. Five hundred turns, 11 levels, and 128 slots were design parameters at the time, without long-term causal validation. [S59]

Gamification can play separate roles: showing users what to do, giving newcomers useful topics, creating shared events, displaying relationship changes, and providing content in stages. If a mechanic only gets users to click through tasks for rewards without improving future dialogue, it may support activity rather than relationships. Compare completion, voluntary return, and negative feedback for the same content with and without compulsory tasks, rather than comparing turns alone.

Doudou obtains ready-made context from the game screen, reducing the burden of deciding what to talk about today. The user's existing main activity supplies topics for the character. But misreading the situation, interrupting a critical operation, or talking too much turns companionship into noise. Quiet periods, responding when summoned, and speaking proactively should be measured separately. Remaining online is a feature of desktop pets; it cannot directly count as retention through deliberate app openings. [S53]

Wanaka's approach through existing friends also has prerequisites: users really have friends willing to play, invitation content has value for recipients, and the first experience is sufficiently smooth. Some Roblox content assumes friends will teach newcomers, but that does not mean a new product can omit onboarding. Editorial inference: first test completing a shared round, accepting invitations, and cohorts that repeatedly play together, then consider instant messaging or additional social layers. One referral event cannot establish that the relationship remains in the product. [S4]

6. Design data, evaluation, and cost for emotional quality together

C.AI's post-training experience suggests that user behavior extends beyond likes and dislikes. Regeneration, editing, and deletion contain preference information, and those preferences shift across user groups and time. The problem is that an event has more than one meaning: regeneration may indicate a boring answer, or a user exploring alternative plots. Without context and reasons, treating every edit as proof that the original answer was wrong turns creative exploration into an incorrectly definite training target. [S43]

EVE tried involving human conversation companions and people with psychology backgrounds, recording exchanges and selecting data to move from assistant-style API replies toward relationship dialogue. However, emotional chain of thought, or emotional CoT, remained a research direction in the program; it cannot be described as a deployed component. Nor can commercial conversation-companion experience be upgraded into evidence of clinical effectiveness. Expensive, carefully selected samples may improve a certain experience while representing only a few people's preferences. They still need checking with target users. [S59]

A testable example set should include memory across days, correction of false memories, character refusals and gentle disagreement, relationship stages that do not jump abruptly, extended non-response, proactive-message frequency, topics the user explicitly does not want to continue, and withdrawal of personal information. Evaluation records should separately assess factual correctness, character fit, emotional appropriateness, and the user's sense of control. A single overall score can hide responses that sound excellent but remember facts incorrectly. Model scoring is useful for screening; it cannot independently replace human experience, user behavior, or review of safety boundaries. This evaluation framework is an editorial synthesis. [S43][S59]

Data collection also affects users and third parties. Me.bot's ideas for always-on devices and low-friction collection do not automatically grant rights to other people's conversations, screens, or photos. Doudou's interpretation of game images on a device likewise needs limits on visibility, storage, and sharing. Users should have separate controls over whether private data is used for long-term memory, training, or other characters. The programs did not provide complete permission implementations, so this article retains these as delivery requirements, not verified capabilities. [S49][S53]

Cost accounting should cover the full relationship service: foreground dialogue, memory extraction, consolidation, retrieval, proactive tasks, images and voice, failed attempts and retries, and content production. EVE's subscriptions plus outfits and stories, Doudou's skins, gifts, and memberships, and FlowGPT's views on in-app added value all show that users pay for different things. Cheap tokens per turn do not establish that lifetime value covers costs, and a company's reported renewal rate for a paying group cannot be treated as data for its whole platform. [S59][S53][S1][S21]

7. Long-term retention and payment: retain counterexamples instead of selecting the most attractive numbers

Several numbers in this collection cannot be directly combined. FlowGPT's roughly 60% entertainment usage was an observation about its platform at the time. Hidecloud's statement that C.AI had roughly 94% renewal lacked a specified sample definition. NieTa used a “421” rule of thumb without a common retention definition. Doudou's six million registrations were neither monthly active users nor paying users. Me.bot's tens of thousands of users and fewer than one hundred continuously trained personal models counted different things. This article does not combine them into an industry average. [S1][S21][S57][S53][S49]

The Doudou episode's strong statement about not looking at retention was later qualified: long-term retention still mattered, while conversation time and turn counts could instead indicate a poor model. Breaking even in the PC business did not mean the whole company was profitable. Complaints that NieTa users had insufficient energy to continue may show a desire to keep going, but whether this converts into willingness to pay, creates resentment, or leads to repeat purchases needs subsequent testing. EVE presented a thorough foundation for everyday interaction after novelty faded, without showing long-term cohort results. A complete design must be separated from demonstrated long-term value. [S53][S57][S59]

An actionable direction for product-team analysis—an editorial inference—is to fix first-use date, acquisition source, language and region, character type, and payment status, then separately measure the first complete experience, return to a fixed character, return to different characters, user-initiated return, relationship continuity in the following week or month, and renewal. Also examine the distribution of bad experiences among frequent users, preventing an improving average from hiding a decline for core users. Content changes, model versions, and event rewards can all change behavior at once. Experiments need these time markers to identify where changes came from.

8. Questions to prioritize, and limits of applicability

1. Do users want to nurture one character or consume more characters? Combine interviews with character switching and return behavior. Do not choose a product direction from a single multiple-choice question.
2. Does memory change the response at the right moment? Record connections across goals, effective corrections, and whether users feel offended together. Merely being able to retrieve information is insufficient for acceptance.
3. Do proactive messages create anticipation or burden? Compare triggers, timing, disable rates, negative feedback, and subsequent voluntary return, rather than only notification clicks.
4. Does story reduce the blank page or obstruct chat? Test the rate of reaching a first meaningful interaction and subsequent self-directed continuation, allowing a choice of lighter and more involved entry points.
5. Is a shared activity more valuable than general conversation? Validate gaming, studying, and creation separately. Do not extrapolate one setting's effect to all companionship.
6. Does payment support the experience, or buy the removal of restrictions? Observe quality after payment, repeat purchases, and regret-related feedback. Short-term conversion cannot replace long-term relationship quality.
7. Can the benefits of personal data be explained to users? Ensure users know what is recorded, why it is used, and how to correct and delete it before evaluating memory depth.

This topic applies to the design of character companions, interactive content, and related AI social products. Not every product needs deep memory, complex planning, multiple characters, or 3D. Simple short-form content may succeed through high-quality supply, and products focused on long relationships may not need open user-generated content. These interviews cover a limited set of companies and guests, with no standardized cross-product experiment. Comparing viewpoints is useful for forming more specific hypotheses and counterexamples, not assembling a standard product guaranteed to succeed.

9. Turn insights across episodes into small experiments that can be reviewed

The following are not experiments already completed by the programs' participants. They are editorial proposals based on eight episodes. Each can be tested in an existing character product, starting with low-risk, reversible versions before deciding whether to expand. Every test needs a record of the target group, version, observation period, failure examples, and exit mechanism. An entertaining one-off demonstration is not a complete conclusion.

Experiment 1: The same memory, used in different ways

Select one preference, one recent goal, and one shared experience that the user has explicitly agreed to save. Construct three response variants: no use of memory, simple repetition of the facts, and connection of the facts to the current question. Have people first assess whether the connection is appropriate, then observe whether users feel understood or need to make corrections. Tests must include counterexamples such as a completed goal, a changed preference, people with identical names, and users who do not want a reminder. If the connected variant is more appealing but makes more mistakes, average satisfaction alone is insufficient; report the severity of false memories as well. EVE and Me.bot provide inspiration for mechanisms. The eventual decision concerns which representation and use create net benefit in this product, not which architecture has the more advanced name. [S59][S49]

Experiment 2: Keep free conversation, while offering shared events with cues

NieTa's findings suggest that some people want to have characters without writing stories. The same character can offer a free-form opening, a brief situational cue, or a prepared event that can be modified after entry. Observe the first complete interaction and later continuation without prompts. To distinguish better content from a reduced burden of expression, hold the character and base-model version constant and let users skip the scenario. After an event, check whether memory accurately carries it forward. If the next interaction still behaves as though nothing happened, event consumption may have succeeded while relationship accumulation has not. Conversely, more active users in the free-form version do not justify removing all guidance: people more practiced at creation may simply have been more willing to enter that group. [S57][S43]

Experiment 3: Evaluate proactive interaction and quiet presence separately

Drawing on Doudou and EVE, first define events that justify a character speaking proactively: something the user shares, a change in a shared activity, or a previously agreed time to check in, rather than forcibly recalling the user after a few idle minutes. Provide visible controls for frequency and quiet periods. Record whether proactive content is accepted, why it interrupts, and whether normal use remains possible after it is disabled. A user not replying does not automatically mean dissatisfaction; replying every time does not necessarily mean enjoyment, either, because politeness or task incentives may be responsible. For a main activity such as gaming, observe whether interruption and ignoring increase. For long-term relationships, also examine whether users independently think of the character and initiate interaction. Both timescales need to be present. [S53][S59]

Experiment 4: Separate creator success from consumer success

FlowGPT and Wanaka both indicate that a team's own production of more characters cannot substitute for platform supply. Recruit a small group of authors willing to maintain characters, and track the process from idea to publication, first useful feedback, revision, and continued creation. On the consumer side, separately record discovery, starting, completing, and returning. If authors use the tool deeply but have no consumers, their debugging time should not count as content popularity. If consumers repeatedly switch works, that can count as successful platform consumption, while leaving open whether the promise of a fixed-relationship product is being delivered. Giving ordinary authors opportunities to be seen can be a product strategy, but whether exposure increases continued creation still requires a separate comparison. [S1][S4]

Experiment 5: Add individual enhancements instead of introducing every modality and multiple characters at once

Hidecloud's personal-photo prototype combined three factors: personal content, the sides characters take in a relationship, and multiple characters. Validation can first add responses to personal material, then the sides characters take, and finally multi-character interaction. Otherwise, even an improved experience will not reveal which factor deserves ongoing expenditure. Images, voice, 3D representations, and proactive tasks each impose latency and maintenance burdens; they should not be collapsed into a single measure of immersion. Users' ability to explain which interaction mattered, their willingness to return to the same experience, and no longer feeling interrupted are closer to product evidence than surprise during a demonstration. A solo self-test can reveal a mechanism, but cannot replace repeated experience among target users. [S21][S59]

These experiments share one principle: the user's sense of control is part of the experience. Clearing memory, switching characters, rejecting proactive topics, or leaving the product should not be treated as hostile behavior that an algorithm must overcome. Long-term value can come from a limited, satisfying content experience or an ongoing relationship. Respect the task being served before discussing monetization and growth, rather than forcing any single behavior into an interpretation of intimacy.

Source index and time context

• S1: Jay / FlowGPT; recorded in November 2023 and published in March 2024. See Original episode.
• S4: Zhang Yang (张阳) / Wanaka; published in February 2026, when the product was in closed testing. See Original episode.
• S21: Zhang Tao (张涛), Hidecloud / product lead at Guangnian Zhiwai; March 2024. See Original episode.
• S43: Ted / former Character.AI post-training specialist; November 2024. See Original episode.
• S49: Tao Fangbo (陶芳波) / Mindverse and Me.bot; August 2024. See Original episode.
• S53: Binson / Doudou and Xinying Suixing; February 2025. See Original episode.
• S57: Zhao Chong (赵充), Joe, and Hu Xiuhan (胡修涵) / AiPPT, ACE Studio, and NieTa; May 2024. See Original episode.
• S59: Tristan / EVE and Singularity Era (《奇点时代》); November 2024. See Original episode.