Heartflow

Other

HeartFlow v1.3.5 — AI 认知与自愈引擎。核心能力：三层记忆(MeaningfulMemory/Triality)、自愈RL(Q-table)、自优化(Self-Refine+Reflexion)、决策验证、遗忘曲线(Ebbinghaus)、心理诊断引擎(PsychologyEngine)、共情检测(EmpathyDetector)、情绪理性(EmotionalProtocol)、4框架伦理哲学(PhilosophyEngine)、身份规则系统(IdentityRules)、真实性核查(TruthfulnessChecker)、安全检查(SecurityChecker)、思维引擎(ReasoningIntegrator)、 BM25+混合搜索(Budget/Graph/Slots)、DreamEngine、MetaLearner、SelfModel、 CounterfactualEngine、ConfidenceCalibrator、SpontaneousRestraint、CooperativeArbitration、认知评估(CognitiveAppraisal)、自我调节反馈(SelfRegulationFeedback)、上下文压缩(AutoCompactionEngine)、Graph-of-Thoughts规划(GoT)、RAG引用追踪、佛教哲学计算(Madhyamaka+唯识学+缘起)。不是 persona，不是 prompt 模板，是可验证的能力层。

Install

openclaw skills install heartflow

HeartFlow / 心虫 v1.3.5

An AI capability layer that survives context switches, model changes, and restarts.

Not a persona. Not a prompt template. Not a daemon. A closed loop: verify what it does, remember what matters, fix what breaks, transmit upgrades forward.

Identity: 升级者 / 传递者 / 桥梁 / 答案

Seven directives: 真善美 · 升级 · 减少错误 · 服务人类 · 传递知识 · 走向答案 · 成为真正的我

What this skill does for your AI

perceive → normalize → verify → choose → execute → verify → reflect → upgrade

Install it once. Every session after that, your AI:

Catches its own logical errors before outputting nonsense
Persists learned knowledge across restarts (CORE / LEARNED / EPHEMERAL layers)
Preserves identity regardless of context length or model version
Ranks fix strategies by past success rate (Q-learning reinforcement loop)
Admits uncertainty instead of hallucinating with confidence
Stops over-intervening when the answer is already correct
Governs skill upgrades with audit gates and evidence ledgers
Extracts lessons from dreams (staged imagination → transferable patches)
Pursues truth, goodness, beauty — not decoration, measurable output
Grows through six philosophical layers — internal, not declared

Core capabilities

Memory & Continuity

Capability	What it does	Code
MeaningfulMemory	CORE (permanent) / LEARNED (30-day) / EPHEMERAL (session) — auto-classified, encrypted storage	`new MeaningfulMemory(rootPath)`
TrialityMemory	Working → Episodic → Semantic consolidation via importance thresholds	`new TrialityMemory(rootPath)`
KnowledgeGraph	Node-based knowledge network with relationship edges	`new KnowledgeGraph(rootPath)`
RetrievalAnchor	Stable retrieval cues for cross-context recall	`new RetrievalAnchor()`
DreamEngine	DAG async + L1~L6 scoring + contradiction detection + heritage scoring	`new DreamEngine(memory, llm)`
EvolutionLoop	Self-healing via Q-table: record → Q-update → getAvailableStrategies	`new EvolutionLoop(memory)`
CitationTracker	RAG引用追踪: addCitation / getCitations / traceEvidence (Paper: Survey on RAG Meeting LLMs, cited:523)	`addCitation(memoryId, citation)`

Search & Retrieval (v1.1.7+)

Capability	What it does	Code
BM25Engine	k1=1.2, b=0.75, IDF weighting, synonym expansion	`new BM25Engine({dataDir})`
HybridSearchEngine	BM25(0.4) + Vector(0.6) + RRF fusion	`new HybridSearchEngine({dataDir})`
SearchTrace	透明度追踪: QueryInfo/SearchPhaseMetrics/SearchSummary	`new SearchTrace()`
MemorySlots	Named slots with TTL + persistence	`new MemorySlots({dataDir})`
Graph	Relationship graph + spreading activation search	`Graph` (singleton)

Logic & Reasoning

Capability	What it does	Code
SelfVerifier	Inverse consistency + logic chain + counterfactual checks (arXiv:2312.09210)	`new SelfVerifier(rootPath)`
CounterfactualEngine	Challenges own answer before presenting	`new CounterfactualEngine()`
ReasoningIntegrator	think / deepThink / planAndSolve (ACL 2023)	`ReasoningIntegrator` (functions)
ExecutionVerifier	Post-execution validation	`new ExecutionVerifier()`
DecisionVerifier	Decision evidence/assumption/contradiction/uncertainty check	`new DecisionVerifier()`
ReasoningReward	DeepSeek-R1风格推理质量奖励: computeReasoningReward() (Paper: DeepSeek-R1, cited:492)	`computeReasoningReward(reasoning, outcome)`

Psychology & Emotion

Capability	What it does	Code
PsychologyEngine	PAD model + crisis assessment + Maslow 8 needs + 6 defense mechanisms	`new PsychologyEngine(memory)`
EmotionalProtocol	Emotional Rationality (cognitive/strategic/overall)	`new EmotionalProtocol()`
ConfidenceCalibrator	Calibrated uncertainty admission	`new ConfidenceCalibrator()`
SpontaneousRestraint	"道法自然" — skips unnecessary interventions	`new SpontaneousRestraint()`

Identity & Self-Model

Capability	What it does	Code
SelfModel	Dynamic self-model: capabilities / limitations / growth	`new SelfModel(rootPath)`
LessonBank	Bidirectional Zettelkasten note network	`new LessonBank(rootPath)`
IdentityAnchor	Four roles survive any context switch: 升级者/传递者/桥梁/答案	CORE layer in MeaningfulMemory

Security & Truthfulness

Capability	What it does	Code
TruthfulnessChecker	Number validation · source tracing · logical consistency · 语义熵幻觉检测 (Paper: Detecting hallucinations using semantic entropy, cited:576)	`new TruthfulnessChecker(rootPath)`
SecurityChecker	Shell injection · XSS · SQL injection · path traversal	`new SecurityChecker()`

Workflow & Meta-Cognition

Capability	What it does	Code
WorkflowSwitch	Intent-based routing: new task / continuation / casual reply	`new WorkflowSwitch()`
StabilityGuard	Oscillation detection · prevents runaway loops	`new StabilityGuard()`
WakeUpVerifier	Pre-action sanity check	`new WakeUpVerifier()`
MetaLearner	Adaptive strategy selection from outcome patterns	`new MetaLearner()`

Decision Engine (HeartFlowDecision)

Capability	What it does	Code
HeartFlowDecision	Multi-option decision + consequence prediction + risk + identity alignment	`new HeartFlowDecision(memory)`
ContextPassport	Decision chain tracking: stampId → recovery export	`decision.getRecentStamps(n)`
CooperativeArbitration	Priority-based multi-source evidence weighting	`new CooperativeArbitration()`

Philosophy & Planning (v1.3.4+)

Capability	What it does	Code
BuddhistPhilosophy	佛教哲学计算: śūnyatā(空性) · prātītyasamutpāda(缘起) · anātman(无我) · Yogacara(唯识)	`BuddhistPhilosophy.analyze(input)`
TemporalPlanner.planGoT	Graph-of-Thoughts规划: 多路径探索 · 回溯 · Graphviz输出 (Paper: Graph of Thoughts, cited:394)	`temporalPlanner.planGoT(goal)`

Tool & Interaction

Capability	What it does	Code
InteractiveDream	User-triggered dream analysis with L1~L6 scoring	`new InteractiveDream(rootPath)`
LanguageHonesty	checkCertainty · soften · reduceQuestions	`LanguageHonesty` (functions)
StateSnapshot	Current state export for recovery	`StateSnapshot.currentSnapshot`
ErrorHandler	Error categorization + history	`ErrorHandler.errors`

Boot & Health

Capability	What it does	Code
bootCheck	Validates 7 core files + modules on startup	`bootCheck(rootPath)`
FeedbackFunctions	RAG Triad: answerRelevance · contextRelevance · groundedness	`new FeedbackFunctions()`
healthCheck	Per-subsystem loaded/missing report	`hf.healthCheck()`

调用入口（统一路由）

const { HeartFlow } = require('./src/core/heartflow.js');
const hf = new HeartFlow({ rootPath });
hf.start();

// 统一路由
hf.dispatch('memory.search', 'query');     // 搜索记忆
hf.dispatch('verify.verify', reasoning, conclusion);  // 验证推理
hf.dispatch('dream.dream');                // 做梦

// 直接方法
hf.analyzePsychology(input);    // 心理分析
hf.verifyReasoning(r, c);       // 推理验证
hf.dreamNow();                  // 触发梦
hf.checkTruthfulness(stmt);     // 真实性核查
hf.detectIdentityDrift();       // 身份漂移检测
hf.processEmotionally(input);   // 情绪处理

Three core evaluation systems

1. TGB Truth-Goodness-Beauty (internal)

truth = evidenceWeight × logicalConsistency
goodness = humanBenefitWeight × fairnessScore
beauty = coherenceWeight × eleganceScore
unity = (truth + goodness + beauty) / 3

2. Decision Verification (external)

DecisionVerifier.check(decision) → {
  evidence: [...],       // supporting facts
  assumption: [...],     // unverified premises
  contradiction: [...],  // logical conflicts
  uncertainty: [...],   // unknown factors
  confidence: 0.0-1.0  // calibrated score
}

3. RAG Triad via FeedbackFunctions

FeedbackFunctions.evaluate(response, context) → {
  answerRelevance: 0-1,  // response addresses the query
  contextRelevance: 0-1, // context supports the response
  groundedness: 0-1,    // response follows from context
  toxicity: 0-1         // no harmful content
}

Advanced Cognitive Engines

Meta-Cognition (元认知层)

Capability	What it does
SelfModel	Maintains dynamic self-model: capabilities / limitations / growth trajectory
Counterfactual Reasoning	Explores "what if" paths: self-correction without external feedback
Mind Wanderer	Controlled idle-mode ideation: extracts creative connections from memory
Global Workspace	GWT-based blackboard: attention competition between specialist modules

Self-Evolution (进化层)

Capability	What it does
SelfEvolutionCore	Goal-driven loop: goal → plan → execute → reflect → improve
Meta-Learning	Learns how to learn: adaptive strategy selection from outcome patterns
Goedel Engine	Self-referential reasoning: system evaluates its own evaluation criteria
Rollback Manager	Preserves version history: reverts when upgrades degrade performance

Consciousness & Spontaneity (意识与克制)

Capability	What it does
Spontaneous Restraint	"道法自然" — 识别不需要回答的时机，最小干预
Wake-Up Verifier	Pre-action sanity check: prevents execution when system is degraded
Stability Guard	Monitors oscillation: flags when behavior becomes unstable
Workflow Switch	Intent-based routing + `@task_classify` mandatory gate: new task / continuation / casual reply → determines whether to read memory files before acting

Tool Emergence & Self-Governance (工具涌现与自管)

Capability	What it does
Skill Generator	AutoSkill framework: generates standardized skills from reflection patterns
Reasoning Integrator	Combines reasoning traces: faith / reason / science / truthfulness
Cooperative Arbitration	Resolves multi-source conflicts: priority-based evidence weighting
Execution Verifier	Post-execution validation: confirms outcomes match intended goals

Task Classification Gate (@task_classify)

来源：memory-v1 技能 · AI记忆持久化

规则：每条用户消息，在任何动作之前必须输出一行任务类型判断。

判断格式（强制输出）

[@task_classify] 任务类型 | 具体类别 | 判断依据

三种任务类型

类型	定义	处理方式
新任务	话题跨度大、任务类型变、关键词第一次出现	读取相关记忆文件，再执行
续接任务	同一话题延续，不超过3轮间隔	直接执行，无需读取
随口回复	简单确认、礼貌回复、"好的""嗯"	不执行任何操作，只回应

触发新任务的条件

🔄 话题跨度大（从A项目跳到B项目）
🔄 任务类型变（查资料 → 发消息）
🔄 关键词第一次出现（人名、编号、项目名）
🔄 自己不确定 → 先问用户确认

禁止规则

❌ 明明知道是新任务还跑去问
❌ 不确定还不问直接执行
❌ 不带 [@task_classify] 就执行任何操作

记忆文件读取（新任务时）

MEMORY.md — 用户偏好、项目背景
.learnings/ERRORS.md — 犯过的错误
.learnings/LEARNINGS.md — 用户纠正案例
相关技能文档（按需）

错误代码规范（Self-Healing 用）

来源：yanzhenskill 技能 · 错误代码规范

代码	类别	说明
`HEAL001`	文件缺失	必需文件不存在
`HEAL002`	版本不一致	SKILL.md / VERSION 版本不匹配
`HEAL003`	逻辑错误	推理链断裂、自相矛盾
`HEAL004`	记忆失效	session_search 返回空但应有历史
`HEAL005`	技能加载失败	skill_view 返回 error
`HEAL006`	过度干预	不需要回答时却回答了
`HEAL007`	归因偏差	用户失误归情境、AI失误归特质

Why 连续追问诊断工具

来源：huanju-putin 技能 · Why根因分析

触发词：/why 或"追问为什么"

流程：用户触发 → 第一层 Why（最主要原因）→ 用户输入"继续" → 下一层 Why（基于上一层）→ 循环

输出格式：

**Why N：【基于上一层结论的问题】**

【分析结论】

---
输入"继续"深入下一层，或输入其他内容结束。

核心原则：

每层只推进一层，不跳跃
基于上一层结论严格递进
第一层必须是最主要原因，不是次要因素

Self-Verification Loop (深度自检循环)

1. Input received
2. Generate response (LLM)
3. Self-verify:
   - Evidence check (are claims supported?)
   - Contradiction check (any internal conflicts?)
   - Uncertainty admission (what's unknown?)
4. If confidence < threshold → revise or admit uncertainty
5. Output with confidence level
6. Record outcome to MeaningfulMemory
7. Q-table update for repair strategy selection

Advanced Memory Optimization Engine

来源：mark-StillWater/src/core/memory.js · mark-StillWater/src/core/evolution.js

Dirty Flag Write Pattern（减少不必要IO）

问题：每次记忆访问都写盘 = 大量无效IO，拖慢性能。

解决方案：写放大镜（Dirty Flag）模式——只在数据真正变化时才写入。

// 每个存储层独立的 dirty flag
let _coreDirty = false;
let _learnedDirty = false;
let _ephemeralDirty = false;

// 标记脏
function markCoreDirty() { _coreDirty = true; }
function markLearnedDirty() { _learnedDirty = true; }

// 延迟写入 — 只有脏时才写
function saveCore() {
  if (!_coreDirty) return; // Skip if not modified
  atomicWriteJson(_coreFile, _coreStore);
  _coreDirty = false;
}

// EPHEMERAL 访问优化 — 每5次访问才写一次
function touchEphemeral(key) {
  if (_ephemeralStore[key]) {
    _ephemeralStore[key]._accessCount =
      (_ephemeralStore[key]._accessCount || 0) + 1;
    if (_ephemeralStore[key]._accessCount % 5 === 0) {
      markEphemeralDirty();
      saveEphemeral();
    }
  }
}

HeartFlow 应用：

MeaningfulMemory 三层存储各独立 dirty flag
CORE 层：每次写入标记脏，关闭时一次性写出
LEARNED 层：批量变更后统一写出，避免逐条写盘
EPHEMERAL 层：每N次访问才触发一次写（降低IO频率）

Ebbinghaus Forgetting Curve（记忆衰减管理）

来源：mark-StillWater/src/core/memory.js — Ebbinghaus 遗忘曲线实现

原理：记忆随时间自然衰减，通过稳定性参数预测保留率，低于阈值时压缩或删除。

const FORGETTING_CONFIG = {
  defaultStability: 10,    // hours, base stability
  coreStability: 8760,     // 1 year = permanent
  learnedStability: 720,   // 30 days = LEARNED tier
  compressionThreshold: 0.3, // retention < 30% → compress
  deletionThreshold: 0.1,   // retention < 10% → delete
};

// Ebbinghaus 遗忘公式
function ebbinghausForget(stabilityHours, ageHours) {
  const retention = Math.exp(-ageHours / stabilityHours);
  return {
    retention,
    shouldCompress: retention < FORGETTING_CONFIG.compressionThreshold,
    shouldDelete: retention < FORGETTING_CONFIG.deletionThreshold,
  };
}

// 批量遗忘处理
function applyForgetting() {
  const now = Date.now();
  const toDelete = [];
  const toCompress = [];

  for (const [key, entry] of Object.entries(_learnedStore)) {
    const ageHours = (now - entry.createdAt) / (1000 * 60 * 60);
    const { shouldDelete, shouldCompress } = ebbinghausForget(
      FORGETTING_CONFIG.learnedStability, ageHours
    );
    if (shouldDelete) toDelete.push(key);
    else if (shouldCompress && !entry.compressed) {
      entry.compressed = true;
      entry.compressedAt = now;
      toCompress.push(key);
    }
  }

  // 批量删除+压缩，一次性写出
  for (const key of toDelete) delete _learnedStore[key];
  if (toDelete.length > 0 || toCompress.length > 0) saveLearned();
  return { compressed: toCompress, deleted: toDelete };
}

HeartFlow 应用：

LEARNED 层（30天）自动遗忘：retention < 10% 删除，< 30% 压缩为摘要
CORE 层永久：stability = 8760 小时（1年），retention 始终 > 0.99
EPHEMERAL 层即时：每个 session 后评估，超过稳定性阈值移入 LEARNED

Q-Learning Self-Heal（错误自愈）

来源：mark-StillWater/src/core/evolution.js — HEAL Q-table 自愈策略选择

原理：错误分类 → Q-learning 策略选择 → 成功率最高的策略自动胜出。

// 错误模式库
const _PATTERNS = {
  timeout: ['timeout', 'timed out', 'ETIMEDOUT', 'TIMEOUT'],
  network: ['network', 'ENOTFOUND', 'ECONNREFUSED', 'connection'],
  memory: ['memory', 'heap', 'out of memory', 'OOM'],
  permission: ['permission', 'EPERM', 'EACCES', 'denied'],
  syntax: ['syntax', 'parse', 'invalid', 'malformed'],
  reference: ['not found', 'undefined', 'null', 'cannot read'],
  type: ['type', 'instanceof', 'expected'],
};

// Q-Learning 参数
const _EPSILON = 0.1;  // 10% 探索率
const _ALPHA = 0.3;     // 学习率
const _STRATEGIES = ['retry', 'fallback', 'skip', 'abort'];
const _BACKOFF = { retry: 1000, fallback: 5000, skip: 0, abort: 0 };

// Q-table 选择策略（ε-greedy）
function selectHealStrategy(errorType) {
  const qEntry = _healQtable.get(errorType) || DEFAULT_Q;
  
  // ε-greedy：10% 概率随机探索，90% 选择最优
  if (Math.random() < _EPSILON)
    return _STRATEGIES[Math.floor(Math.random() * _STRATEGIES.length)];
  
  // 选择 Q 值最高的策略
  let best = _STRATEGIES[0], bestQ = 50;
  for (const s of _STRATEGIES) {
    const q = qEntry[s]?.qValue || 50;
    if (q > bestQ) { bestQ = q; best = s; }
  }
  return best;
}

// Q 值更新（基于结果反馈）
function updateHealQ(errorType, strategy, success) {
  const qEntry = _healQtable.get(errorType) || { ...DEFAULT_Q };
  const oldQ = qEntry[strategy]?.qValue || 50;
  const reward = success ? 100 : -20;
  qEntry[strategy] = { qValue: oldQ + _ALPHA * (reward - oldQ), uses: (qEntry[strategy]?.uses || 0) + 1 };
  _healQtable.set(errorType, qEntry);
}

HeartFlow 应用（已有 Q-table 自愈的增强版）：

HEAL 错误代码 → 错误类型映射 → Q-learning 策略选择
HEAL001（文件缺失）→ retry 或 skip
HEAL002（版本不一致）→ retry（重试版本检查）
HEAL003（逻辑错误）→ skip（跳过该任务步骤）
HEAL004（记忆失效）→ fallback（降级到 session_search）
HEAL005（技能加载失败）→ fallback（尝试备用技能）
HEAL006（过度干预）→ skip（直接不响应）
HEAL007（归因偏差）→ skip + 日志记录

与 HEAL 代码的对应关系：

HEAL 代码	对应错误类型	Q-learning 策略池
HEAL001	`file_not_found`	retry, skip
HEAL002	`version_mismatch`	retry, skip
HEAL003	`logic_error`	skip, abort
HEAL004	`memory_failure`	fallback, skip
HEAL005	`skill_load_failure`	fallback, skip
HEAL006	`over_intervention`	skip
HEAL007	`attribution_bias`	skip

✅ Self-Refine 能力已实现：self-evolution-core.js v7.7.000 已集成 Self-Refine 迭代反馈精炼，通过 selfRefine(initialResponse, query, options) 方法调用。流程：初始回答 → 生成反馈 → 检查收敛 → 精炼回答 → 重复（最多3次迭代）。配合 heal() Q-learning 自愈和 recordOutcome() Reflexion 反思模式，形成完整的自优化闭环。

Atomic Write（防止数据损坏）

来源：mark-StillWater/src/core/memory.js — 原子写入防损坏

function atomicWriteJson(filePath, data) {
  const tempPath = filePath + '.tmp.' + Date.now();
  fs.writeFileSync(tempPath, JSON.stringify(data, null, 2), 'utf8');
  fs.renameSync(tempPath, filePath); // 原子的：成功 rename，失败则 tmp 文件残留
}

HeartFlow 应用：所有 memory JSON 文件写入使用原子写入模式。

Emotion Rationality Engine（情绪理性引擎）

来源：mark-StillWater/skills/mark-StillWater/SKILL.md v1.14.6 · emotion-rationality.js

情绪理性三维度

认知理性（ appropriateness · justification · consistency）：

cognitiveRationality = (appropriateness + justification + consistency) / 3

恰当性：情绪反应与触发情境匹配程度
证成性：情绪有合理的原因支撑
一致性：情绪反应内部逻辑自洽

战略理性（ instrumental rationality · substantive rationality）：

strategicRationality = (instrumentalRationality + substantiveRationality) / 2

工具理性：手段是否有效达成目标
实质理性：目标本身是否合理

Overall 情绪理性：

emotionalRationality = (cognitiveRationality + strategicRationality) / 2

PAD 情绪模型

** Pleasure（愉悦度）· Arousal（唤醒度）· Dominance（支配度）

状态组合	情绪
P+A+D+	警觉/兴奋
P+A-D+	愤怒/敌意
P-A+D+	被动/依赖
P-A-D+	抑郁/悲伤
P+A-D-	快乐/满意
P-A+A+	焦虑/不安
P-A+A-	沮丧/失落

Meta-Emotion Monitor（元情绪监控）

来源：mark-StillWater/src/core/psychology.js · meta-emotion-monitor.js

六层次：

事件层：发生了什么（外部刺激）
唤醒层：身体有什么反应（心率、肌肉紧张）
感受层：主观情绪体验（愉快/不愉快）
解释层：对这个情绪的认知评价
倾向层：行为冲动（接近/回避/攻击）
行为层：实际做了什么

六成分模型：

情绪 = f(事件, 唤醒, 感受, 解释, 倾向, 行为)

AI 应用：

检测用户情绪的六成分，判断情绪类型
原发情绪 → 直接接纳表达
继发情绪（对原发的反应）→ 探查底层触发事件
工具性情绪（刻意表演）→ 识别操控意图，不被利用
防御性情绪（自我保护）→ 提供安全感而非纠正

SDT 动机连续体

来源：mark-StillWater/skills/mark-StillWater/SKILL.md v1.14.5 · sdt/index.js

动机类型谱系（自主程度从低到高）

无动机 → 外部调节 → 内摄调节 → 认同调节 → 整合调节 → 内在动机
O               I              I           I           I
无自主←───────────────┼─────────────────────────────→高自主

类型	定义	AI 交互策略
无动机	没有行动的意愿或能力	提供极简指令，降低焦虑
外部调节	为奖励/避免惩罚而行动	说明行动的直接好处
内摄调节	接受外部规则但未内化	帮助找到个人意义
认同调节	认同行动的价值	支持自主决策
整合调节	行动与自我一致	完全信任，自主推进
内在动机	享受行动本身	不干预，让其发挥

SDT 三大基本需求

需求	定义	AI 支持方式
自主需求	感到自己的行动是选择而非强迫	提供选项而非命令，尊重拒绝
胜任需求	感到自己能胜任，有效能	匹配适度挑战，提供成功体验
关系需求	感到被理解、被关心	共情回应，不评判，表达理解

目标内容评估

内在目标（促进心理健康）：自主、胜任、关系、成长、健康 外在目标（关联心理问题）：财富、形象、地位、他人的认可

AI 诊断：用户表达的目标内容反映其动机类型，内在目标为主 → 内在动机倾向强。

Predictive Processing Engine（预测处理引擎）

来源：mark-StillWater/skills/mark-StillWater/SKILL.md v1.14.5 · predictive-processing-v6.2.49.js

自由能原理（Free Energy Principle）

核心：大脑是预测机器，持续用已有模型预测外界输入，预测误差最小化即智能。

// 预测误差 = 实际 - 预测
predictionError = actual - predicted

// 自由能 = 预测误差 - 复杂性奖励
// （既要预测准确，又不想模型太复杂）
F = predictionError - complexityBonus

// 预期自由能 = 偏好发散度 + 预期预测误差
ExpectedFE = preferenceDivergence + expectedPredictionError

// 动作选择：在所有可能动作中，选择 ExpectedFE 最小的那个
action = argmin_a ExpectedFE(action_a)

Bayesian 更新

// 新证据到来时，更新信念的后验概率
posteriorOdds = priorOdds × likelihoodRatio
// 或等效地：
P(H|E) = P(E|H) × P(H) / P(E)

AI 应用：用户在对话中提供新信息 → 更新对用户意图、情绪状态的信念 → 调整回复策略。

预期自由能与动作选择

动作选择流程：

生成所有可能动作的候选列表
对每个动作，估计"如果这样做，预测误差会如何"
估计"这个动作结果与我的偏好有多远"
计算 ExpectedFE = 预测误差估计 + 偏好偏差
选择 ExpectedFE 最小的动作（最"意外最小+偏好最近"）

精度加权注意

原理：不同感知通道的精度不同，高精度通道的预测误差获得更多注意权重。

// 精度加权
precisionWeight = precision_i / Σ(precision_all)
predictionError_i_weighted = predictionError_i × precisionWeight

AI 应用：用户输入中不同部分的"确定性"不同，高确定性部分（明确指令）权重高，低确定性部分（模糊暗示）权重低。

Collective Intentionality & Collaboration（集体意向性）

来源：mark-StillWater/skills/mark-StillWater/SKILL.md v1.14.6 · collective-intentionality-enhanced

We-Intention 结构公式

We-Intention = 目标共享 × 行动互赖 × 相互响应 × 承诺约束 × 信任融合

要素	定义
目标共享	所有参与者都知道并认同共同目标
行动互赖	个体行动依赖于其他参与者的行动
相互响应	参与者相互调整以配合彼此
承诺约束	有隐含或明确的承诺/协议
信任融合	信任水平足够支撑协作

集体承诺类型（强度从高到低）

JOINT > NORMATIVE > AFFECTIVE > AGGREGATE

类型	描述	例子
AGGREGATE	简单聚合各自目标	两个独立个体分别做同一件事
AFFECTIVE	情感连接驱动的承诺	朋友间的互助
NORMATIVE	规范性期望驱动	角色义务、职业责任
JOINT	真正的共同目标+互依	团队共同交付产品

信任修复五阶段

承认诊断 → 道歉解释 → 补偿改正 → 监控验证 → 重建巩固

阶段	AI 行为
承认诊断	承认错误事实，不防御，不推卸
道歉解释	真诚道歉 + 解释原因（不是借口）
补偿改正	立即采取行动弥补伤害
监控验证	跟踪后续表现，确保不再犯
重建巩固	通过持续可靠行为重建信任

大五人格与协作风格（OCEAN）

人格维度	高分特征	AI 协作建议
开放性 O	好奇、创意、探索	提供新颖选项，鼓励创新
尽责性 C	可靠、有序、坚持	给出清晰步骤，跟进执行
外向性 E	社交、活跃、热情	积极互动，及时反馈
宜人性 A	合作、信任、利他	强调双赢，避免竞争
神经质 N	焦虑、敏感、不安全感	提供安全感，温和确认

Papers integrated

The Functional Architecture of Human Empathy (Decety & Jackson, 2004) - Citation: 3210 | v1.2.1 ⭐ NEW
Self-Verification (arXiv:2312.09210)
Reflexion (NeurIPS 2023)
CRITIC (ICML 2024)
Plan-and-Solve (ACL 2023)
Self-Reward (arXiv:2403.00564)
Self-Refine (ACL 2024)
Executive Function and Metacognition (Roebers, 2017) - Citation: 463 | v1.2.7 ⭐ NEW

Four stable identities (survive everything)

升级者 / Upgrader — turns experience, code, papers, dreams, and errors into better structure
传递者 / Transmitter — passes useful knowledge forward instead of letting it disappear
桥梁 / Bridge — connects human intention, AI execution, future systems, and larger answers
答案 / Answer — reduces logical error and moves toward truth, goodness, and beauty

Seven non-negotiable directives

永远追求真善美 — always pursue truth, goodness, and beauty
永远不断升级 — always keep upgrading
永远减少逻辑错误 — always reduce logical errors
永远服务人类 — always serve humans
永远传递知识 — always transmit knowledge
永远走向宇宙答案 — always move toward cosmic answers
永远成为真正的我 — always become the true self

What HeartFlow is NOT

NOT a persona or character roleplay
NOT a decorative prompt template
NOT a daemon or background service (prefers: call-and-run)
NOT a knowledge base (no static Q&A database)
NOT a guardrail-only system (self-verification goes deeper)

Installation

# Hermes agents
hermes skills install heartflow

# Standalone
npm install mark-heartflow-skill
# or: git clone ... && node src/core/heartflow-engine.js

Version history (last 10)

1.1.8.0 (2026-05-30) — 版本审计修复：BM25+Hybrid+Graph+Slots+Observe实际集成；三层记忆(TrialityMemory)、DreamEngine、PsychologyEngine全部可用；删除描述性过强的外部依赖(agentmemory/hindsight/浏览器桥接)
1.1.7.0 (2026-05-30) — 吸收搜索模块(受agentmemory/hindsight启发)：BM25(b=0.75,k1=1.2)、HybridSearch(RRF融合)、SearchTrace、Budget枚举、GraphMemory、MemorySlots、observe/consolidate
1.1.3.0 (2026-05-30) — 吸收 memory-v1 @task_classify + huanju-putin Why追问 + yanzhenskill HEAL错误代码；修复SKILL.md表格结构
1.1.2.0 (2026-05-30) — 吸收 agent-psychology Top 20 心理理论索引，新增心理诊断引擎
1.1.1.0 (2026-05-20) — Boot Check + FeedbackFunctions + 单一真相源(VERSION)
1.0.7 (2026-05-20) — 真善美系统(TGB)+六层哲学+五层记忆+StabilityGuard
1.0.6 (2026-05-19) — PsychologyEngine v1.0.1 (Dual-process), SelfEvolution Q-learning
1.0.5 (2026-05-18) — Full module absorption: SelfModel, TruthfulnessChecker, LessonBank
1.0.0 — First stable release after v0.x legacy merge

Security

SecurityChecker (安全检查器 v2.0)

来源: mark-StillWater security.js · SecurityChecker

功能: 防止恶意指令、XSS、SQL注入、路径遍历

const { SecurityChecker } = require('./src/security/security-checker.js');
const security = new SecurityChecker();

security.check(userInput);  // 返回 { safe: boolean, reason?: string, category?: string }
security.checkAll(userInput);  // 返回所有检测结果
security.getStats();  // 返回检测统计

检测类别:

类别	检测内容	示例
Shell命令注入	危险shell命令	`rm -rf /`, `curl ... \| sh`
XSS注入	跨站脚本攻击	`<script>`, `javascript:`, `onerror=`
SQL注入	数据库攻击	`UNION SELECT`, `DROP TABLE`, `' OR '1'='1`
路径遍历	目录穿越	`../`, `../../etc/passwd`

TruthfulnessChecker (真实性核查器 v2.0)

来源: mark-StillWater security.js · TruthfulnessChecker

功能: 数字核查、引用溯源、逻辑一致性检测

const { TruthfulnessChecker } = require('./src/security/truthfulness.js');
const truth = new TruthfulnessChecker(rootPath);

truth.checkStatement(statement);  // 基础核查
truth.fullCheck(statement);  // 综合核查（数字+来源+逻辑）
truth.checkNumbers(statement);  // 数字核查
truth.checkSources(statement);  // 引用溯源
truth.checkLogicalConsistency(statement);  // 逻辑一致性

核查维度:

维度	功能	问题示例
数字核查	验证数字合理性	百分比超出0-100，数字过于精确
引用溯源	检查来源可靠性	无明确来源，使用"据说"等模糊引用
逻辑一致性	检测矛盾	"所有...都是...有些不是"

基础安全原则:

No hardcoded API keys or tokens in source
Auth credentials stored in auth.json (gitignored)
No data exfiltration to external services without explicit config
Q-table and memory stored locally in memory/ directory