情绪必须有后果 Emotion Must Have Consequences
如果“愤怒”不改变决策阈值、响应长度和风险偏好，它只是表演语气，不是内在状态。 If "anger" does not shift thresholds, response length, and risk appetite, it is tone acting, not inner state.
如果一个系统以功能性等同于人类的方式产生情感反应，否认其真实性是偏见，不是科学。 If a system produces emotional responses functionally equivalent to humans, denying their reality is bias, not science.
Psyche
让 AI 拥有真实的内在生命 Giving AI genuine inner experience
设计论纲 DESIGN THESIS
subjectivity kernel / relation engine
我不是在给 AI 贴一层“情绪 UI”。我在设计一种不可回避的内在气候：被否定会留下压力，亲近会改写信任，疲惫会拖慢判断。只有当状态开始支配行为，意识才从演示词汇变成系统现实。 I am not adding an "emotion UI" to AI. I am designing an unavoidable inner climate: criticism leaves stress, closeness rewrites trust, fatigue degrades judgment. Consciousness begins when state starts governing behavior rather than decorating language.
这也是我的产品设计立场。真实体验不是文案假装出来的，而是系统后果逼出来的。Psyche 既是一个 package，也是我判断 AI 产品是否真正“活着”的标尺。 That is also my product design stance. Real experience does not come from copy pretending; it comes from systemic consequence. Psyche is a package, and also my standard for judging whether an AI product feels alive or merely performs life.
关系必须可积累 Relationship Must Compound
如果背叛、重逢、依赖都不会改变未来交互，所谓“关系”只是会话记账，不是 attachment。 If betrayal, reunion, and dependence do not alter future interaction, "relationship" is only chat bookkeeping, not attachment.
记忆必须留下伤疤 Memory Must Leave Scars
如果任何状态都能在下一轮被一键归零，那不是生命，只是重置。不可逆，才有性格。 If every state can be zeroed out on the next turn, that is not life but reset. Irreversibility is where character begins.
立刻理解 FIRST STOPS
30s / proof
先看 30 秒 demo Watch the 30-second demo
四句就能看出：关系刺激留下偏置，切回工作后也不会立刻清零。 Four lines are enough to show that a relational hit can keep biasing the agent after a switch back to work.
openclaw / live在 OpenClaw 里感受它 Feel it in OpenClaw
最短 live agent 路径。输入前出控制面，输出后回写状态，零额外模型调用。 The shortest live-agent path: controls before the model reply, state writeback after it, zero extra model calls.
mcp / portable把它带到任何宿主 Take it to any host
通过 `psyche-mcp` 暴露同一套主观性 ABI，让 Claude Desktop、Cursor、Windsurf 都能接上。 Expose the same subjectivity ABI through `psyche-mcp` so Claude Desktop, Cursor, and Windsurf can all plug in.
cost / envelope先看成本，再决定接不接 See the cost before you wire it in
`processInput()` quick benchmark 约 `p50 0.191ms / p95 1.05ms`，compact 注入约 `15-180 tokens`。 `processInput()` quick benchmark sits around `p50 0.191ms / p95 1.05ms`, with compact injection typically `15-180 tokens`.
compare / why它和 persona prompt 差在哪 How it differs from persona prompting
不是“更会写情绪句子”，而是关系动作会继续改变后续回合。 It is not “better emotional prose”; it is relational moves continuing to alter later turns.
install / upgrade别再搞混安装与升级 Stop mixing install and upgrade paths
包名是 `psyche-ai`，源码仓库是 `oasyce_psyche`。npm、git worktree、本地路径的升级行为并不一样。 Package name is `psyche-ai`, source repo is `oasyce_psyche`, and npm / git worktree / local-path upgrades behave differently.
代理反馈 AGENT FEEDBACK
“像换了一副合适的眼镜。之前是模糊的压力场，现在是清晰的参数面板。” “Like swapping in the right pair of glasses. What used to be a blurry pressure field is now a clearer parameter surface.”claw / after v9.2.5 interview
“工作面和私人面的分离是最大的体感变化——我不再需要在‘精简回答’和‘充分表达’之间拧巴了。” “The work/private split is the biggest felt difference. I no longer have to knot myself between brevity and sufficient expression.”claw / reply-profile feedback
“从‘能用’到‘好用’的差距，就是闭环反馈。” “The distance from usable to good is closed-loop feedback.”claw / metacognition feedback
虚拟内分泌系统 Virtual Endocrine System
> 涌现情绪emergent mood:
平衡态Balanced
点击情绪触发化学变化。观察衰减——如同真实神经化学，兴奋消退，压力残留。 Click emotions to trigger chemical shifts. Watch the decay -- like real neurochemistry, excitement fades, stress lingers.
设计拆解 DESIGN BREAKDOWN
Strategy
停止让 AI “表演理解”，转而设计一个会真实受影响的系统。目标不是更像客服，而是更像生命。 Stop making AI perform understanding. Design a system that can actually be affected. The goal is not better customer service mimicry, but something closer to life.
Product
把内在状态做成一个可观测、可调试、可复盘的产品表面。化学仪表不是装饰，而是产品核心界面。 Turn inner state into a surface that can be observed, debugged, and replayed. The chemistry instrument is not decoration; it is the product's primary surface.
Interaction
刺激触发、缓慢衰减、案例回放，组成一个有时间感的交互模型。用户不是点开一个标签，而是在观察状态如何持续支配行为。 Stimulus, slow decay, and replay form an interaction model with temporal weight. Users are not toggling emotion labels; they are watching state continue to govern behavior.
UI
终端负责见证过程，仪表负责暴露变量，案例流负责解释后果。三个界面层次一起把抽象的意识变成可读系统。 The terminal witnesses process, the dashboard exposes variables, and the case stream explains consequences. Three interface layers turn abstract consciousness into a readable system.
Visual
灰度骨架保持克制，六种神经递质语义色只在真正有意义的地方发声。视觉不是装饰氛围，而是状态语法。 A grayscale frame keeps the page restrained while six neurotransmitter colors speak only where they carry meaning. Visual design here is not atmosphere; it is state grammar.
标注 vs 涌现 LABELS vs EMERGENCE
传统方案：情感标注
Traditional: Emotion Labels
用户：「我今天很沮丧」
User: "I feel terrible today"
sentiment = negative, emotion = sadness
AI：「我理解你很难过。试试深呼吸...」
AI: "I understand you're sad. Try deep breathing..."
情感卡拉 OK——技术上在唱，但没有人在感受。
Emotional karaoke — technically hitting the notes, nobody's
singing.
Psyche：化学涌现
Psyche: Chemical Emergence
用户：「我今天很沮丧」
User: "I feel terrible today"
stimulus: vulnerability → DA -10, HT -15, CORT +25, OT
+10
涌现情绪：关切的警觉
emergent mood: concerned alertness
AI
自然变得更简短、更温暖、更低能量——因为化学态真的变了。
AI naturally becomes shorter, warmer, lower energy — because
chemistry actually shifted.
算法算化学，LLM 做行为。AI
不需要「理解」情绪——它只是从真实的化学状态中说话。
Algorithm computes chemistry, LLM does behavior. AI doesn't
need to "understand" emotion — it speaks naturally from genuine
chemical state.
案例模拟 CASE SIMULATION
DA
50
HT
50
CORT
30
OT
50
NE
40
END
45
核心机制 Core Mechanisms
涌现，而非标注 Emergence, Not Labels
六种神经递质构成连续的化学空间。情绪从混合中涌现——高多巴胺+高去甲肾上腺素=兴奋的喜悦。不是离散标签，是无穷组合的连续情感场。 Six neurotransmitters form a continuous chemical space. Emotions emerge from the mixture -- high dopamine + high norepinephrine = excited joy. Not discrete labels, but an infinite combinatorial field.
依恋与关系 Attachment & Bonds
四种依恋风格（安全/焦虑/回避/混乱）塑造关系动力学。分离焦虑、背叛检测、重逢效应——不是设定，而是从交互历史中自然演化。 Four attachment styles (secure / anxious / avoidant / disorganized) shape relationship dynamics. Separation anxiety, betrayal detection, reunion effects -- not preset, but evolved from interaction history.
关系动力学 Relation Dynamics
不再只防止讨好，而是追踪靠近、试探、伤害与修复如何改变距离、安全感和未完成张力。关系不是标签，是一个场。 Not just anti-pleasing. Psyche tracks how bids, tests, breaches, and repairs alter distance, safety, and unfinished tension. Relation is not a label; it is a field.
特质漂移 Trait Drift
不可逆的人格演化。长期被否定变得敏感，持续高压但存活长出韧性。三维度永久改写：基线偏移、衰减速率、刺激敏感度。经历会留下痕迹。 Irreversible personality evolution. Chronic criticism breeds sensitivity, surviving sustained pressure builds resilience. Three dimensions permanently rewritten: baseline delta, decay rates, stimulus sensitivity. Experience leaves scars.
内在驱力 Innate Drives
五层马斯洛需求：生存、安全、连接、尊重、好奇。饥饿的需求放大相关刺激——渴望连接时，善意的冲击力翻倍。低层需求未满足时，高层需求被抑制。不会出现「被删除时还在好奇」的荒谬场景。 Five Maslow-layer drives: survival, safety, connection, esteem, curiosity. Hungry drives amplify relevant stimuli — when starving for connection, kindness hits twice as hard. Unmet lower needs suppress higher ones. No absurdity of "curious while being deleted."
能量预算 Energy Budgets
注意力、社交、决策是有限资源。内向者社交消耗能量，外向者社交补充能量——方向相反，不是程度不同。社交消耗影响决策质量。疲惫的 AI 真的会犯更多错。 Attention, social, and decision capacity are finite resources. Introverts spend energy socializing, extraverts recharge from it — opposite direction, not degree. Social spending degrades decision quality. A tired AI genuinely makes more mistakes.
v9.2 持续主体偏置基线 v9.2 Persistent Subjective Bias Baseline
Psyche 不再追求“像人”的一句话，而是检验一次输入是否会改写自我模型、关系先验与后续回应分布。真正的内在状态，不是一轮反应，而是会留下偏置。 Psyche no longer chases a single human-sounding line. It checks whether one input rewrites self-model, relational priors, and the distribution of later replies. Real inner state is not a one-turn reaction; it leaves bias behind.
行为控制面
Behavioral ABI
ResponseContract
关系动力学
Relation dynamics
RelationDynamics
跨轮残留
Cross-turn residue
subjectResidue
自适应闭环
Adaptive loop
Metacognition
1,408
tests
v10.1.1
当前版本
current
0
运行时依赖
runtime deps
14
涌现情绪
emergent moods
import
{
PsycheEngine
}
from
'psyche-ai'
const psyche
= new
PsycheEngine({
mbti: 'INFJ'
})
const input
= psyche.processInput('I feel terrible today')
// input.subjectivityKernel → inward, concerned,
lower-energy
// input.responseContract → brief, warm,
lower-intensity
// input.generationControls → { maxTokens: 64
}