这可能才是 AI 陪伴真正该有的样子｜对谈刷屏产品 EVE 创始人 Tristan

Summary

This 42章经 episode interviews Tristan, founder of Natural Selection / 自然选择, about EVE as a long-term AI companion product. Tristan argues that EVE is not primarily Character AI-style roleplay or interactive fiction; it is closer to an AI long-distance partner that uses AI Companion Active Memory, real-world temporal awareness, proactive multimodal interaction, game progression, and emotional post-training to build relationship continuity. The source deepens the wiki’s AI Friend Products branch by showing how a romance-game team turns character design, 3D production, relationship progression, and human companion-chat data into an AI-native consumer product.

Key Claims

• Tristan defines companionship as having a “super-aligned” partner who understands a user’s history, values, goals, and current world rather than merely sustaining chat time.
• EVE rejects the idea that high session duration alone proves companionship; Character AI-style products may produce deep content consumption without creating a stable relationship object.
• AI Companion Active Memory is the core technical bet: EVE uses roughly 128 memory slots, always-on reflection, slot classification, memory merge, and long-lived key context so user goals and preferences can resurface without keyword prompting.
• The product accepts slightly slower replies when stronger planning, memory recall, topic generation, search, or deeper model routing can make the answer feel more thoughtful.
• EVE uses game loops as relationship cold start: first-episode onboarding, tasks, relationship levels, unlocks, 3D scenes, outfits, and interactive content help users talk enough for the companion to learn them.
• Real-world temporal awareness matters because long-term companionship needs new material from the outside world; EVE can discuss hot topics, search when needed, and push timely messages or memes.
• Natural Selection / 自然选择 treats romance-game experience as an advantage in character design, emotional value, 3D pipelines, onboarding, rewards, and monetization, while still keeping daily chat and companionship as the product core.
• The source frames emotional post-training as necessary after base model APIs made the product feel too assistant-like; the team collected human companion-chat data and hired emotional-interaction designers, including psychology backgrounds and top human chat companions.
• EVE’s “independent personality” design includes negative feedback: disrespectful user behavior can reduce trust or lead the character to withdraw, supporting the feeling that the companion is not merely a tool.
• EVE’s economics combine subscription limits with game-like purchases such as outfits, card stories, and 3D interactions, while Tristan argues the product must keep cost below LTV and benefit from falling token prices.

Key Quotes

“超级对齐的灵魂伴侣” — Tristan on the first principle of companionship.

“RAG 是被动记忆” — Tristan on why ordinary retrieval is not enough for a companion.

“文字和灵魂仍然是最核心的” — Tristan on the role of 3D versus conversation quality.

Connections

• Tristan, Natural Selection / 自然选择, and EVE — founder, company, and companion product at the center of the episode.
• AI Friend Products, AI Companion Active Memory, Emotional Interaction Models, and Persistent Agent Memory — main companion and memory concepts extended by the source.
• Proactive Agents, Context Engineering, and AI Native Product Design — product architecture behind timely messages, topic creation, and context-aware replies.
• AI Startup Unit Economics, Product Led Willingness To Pay, and AI Interactive Content Platforms — commercialization themes around high-cost chat, subscription limits, and game-like paid content.
• Character AI, AI Otome Games, and AI Interactive Entertainment — comparison points for roleplay, companion games, and interactive content.
• Elys, AI Social Networks, and Cyber Avatars — adjacent Natural Selection branch from the later Tristan source, where AI becomes a bridge to human connection rather than the main companion object.

Contradictions

• No direct contradiction with existing wiki content. The source strengthens the existing EVE page by adding implementation detail that the earlier Elys interview only summarized.
• It qualifies broad AI-companion engagement claims by arguing that time spent can indicate content consumption rather than companionship, especially for Character AI-style roleplay products.
• It also qualifies pure-model explanations of AI companionship: Tristan presents the product as a system of memory, post-training, game design, 3D interaction, world awareness, and relationship-state mechanics rather than a model wrapper alone.