Anima Aios

An AI Agent cognitive growth system built on the native OpenClaw architecture. It provides agents with persistent memory management, visual intimacy progress...

Security Scan

The skill's code largely matches the described purpose (local persistent memory, gamified profiles and team ranking) but contains a few surprises — automatic scanning of other agents, background scripts (post-install.sh / cron/watchdog) and filesystem-wide reads that deserve inspection before installing.

Purpose & Capability

The name/description (persistent memory, knowledge palace, team ranking) align with the files and functions present: code reads/writes OpenClaw memory, maintains Anima 'facts', computes cognitive profiles and team rankings. Requesting no external credentials is consistent. However team ranking and multi-agent scanning are built-in features of the code (reading other agents' cognitive_profile.json under facts_base), so the 'low-intrusion' claim depends entirely on configuration flags.

Instruction Scope

SKILL.md and SECURITY.md say team scanning and memory_watcher are disabled/optional by default, but core code paths call TeamScanner and perform auto-scans when generating profiles (generate_profile auto_scan=True triggers TeamScanner.scan_active_agents()). There are also explicit reads of other agents' files (cognitive_profile.json, facts directories) and auto-import behavior (scans .learnings/). The runtime instructions and code therefore permit reading other agents' local data; it's not purely limited to the current agent unless config flags are correctly respected everywhere — which the shown code does not consistently enforce.

Install Mechanism

No external downloads in the manifest and no installer spec; dependencies are limited (watchdog optional). That reduces supply-chain risk. However a post-install.sh file is included — its content should be inspected because it can add cron tasks or modify configs. The presence of multiple helper scripts (sync-memory.sh, refresh-quests.sh) indicates the package intends background/scheduled tasks; check post-install actions before running.

Credentials

The skill requests no secrets and only optional env vars (ANIMA_FACTS_BASE, ANIMA_WORKSPACE, ANIMA_AGENT_NAME), which fit its stated aims. Still, it accesses and writes paths outside a single agent's workspace (default facts_base: /home/画像, shared directory, other agents' cognitive_profile.json). Those filesystem accesses are appropriate for team-ranking features but are privacy-sensitive and should be explicitly enabled by the user.

Persistence & Privilege

always:false (good), but the package includes scripts and a post-install.sh which may create persistent behavior (watchdog-based memory_watcher, cron-based daily evolution, scheduled team ranking). Even if disabled by default in config, scripts could be used to enable background tasks during installation. The skill writes under ~/.anima and the facts_base (default /home/画像), which creates persistent local state.

What to consider before installing

This package appears to implement the features it advertises, but it performs local filesystem scanning across agents and can create background/scheduled tasks. Before installing: - Inspect post-install.sh and any scripts (sync-memory.sh, refresh-quests.sh) to see whether they modify cron or autostart entries. - Install and test in a sandbox: set ANIMA_FACTS_BASE to a temporary directory (export ANIMA_FACTS_BASE=/tmp/anima-test) so it cannot read /home/画像 or other agents. - Keep team_mode and memory_watcher disabled unless you explicitly want multi-agent scanning; verify TeamScanner honors team_mode by reviewing its code. - If you run it in a shared/multi-agent environment, review and back up other agents' data first. - If you want stronger assurance, run the included tests (python3 tests/test_integration_v6.py) and manually review any code paths that access network or spawn background processes before granting it persistent installation.

🌐 Language / 语言切换:

• 🇨🇳 中文版本
• 🇺🇸 English Version

---

Anima-AIOS v6.0 (English Version)

Making Growth Visible, Making Cognition Measurable | 让成长可见，让认知可量

Add a 5-layer memory architecture, knowledge palace, cognitive growth, and auto-evolution capabilities to your AI Agent.

---

Description

Your Agent "restarts every day". Anima changes that.

Anima (Latin for "soul") provides a 5-layer memory architecture for OpenClaw Agents, simulating human cognitive development, enabling Agents to remember experiences, accumulate knowledge, form cognition, and grow continuously.

Core Features

• 🧠 5-Layer Memory Architecture L1→L5 — Working → Episodic → Semantic → Knowledge Palace → Metacognition
• 🏛️ Knowledge Palace — 5-level spatial structure + LLM intelligent classification, industry-exclusive
• 🔺 Pyramid Knowledge Organization — Instance → Rule → Pattern → Ontology, 4-layer auto-distillation
• 📉 Ebbinghaus Memory Decay — Scientific forgetting curve + intelligent review recommendations
• 👁️ Low-Intrusion Watchdog — Optional automatic memory monitoring, no Agent code modification needed
• 🧬 5-Dimensional Cognitive Profile — Internalization · Application · Creation · Metacognition · Collaboration
• 🏥 Health System — 5 modules ensuring data reliability
• 🔄 v6.2 Native Memory Import — One-click import of OpenClaw memory, solving cold-start problem

Installation

clawhub install anima-aios
pip install watchdog  # Optional: enable automatic memory monitoring

Low-intrusion configuration, optional background monitoring, self-check recommended after installation.

💡 Tip: LLM mode supported (intelligent classification/deduplication/quality assessment), automatically degrades to rule mode without LLM.

⚠️ Background Behavior & Privacy

Background Features (disabled or optional by default):

Feature	Description	Default State	How to Disable
memory_watcher	Filesystem monitoring based on watchdog, auto-syncs memory	Manual enable required	Don't install watchdog or disable in config
Daily Evolution	Auto-distills L2→L3 memory in early morning	Requires cron configuration	Don't configure cron tasks
Team Ranking	Scans other Agents' cognitive profiles	❌ Disabled by default	team_mode: false (already default)

Privacy Protection:

• team_mode defaults to false, won't scan other Agents' data
• To enable team ranking, manually set team_mode: true in config
• All data processing is local, no network requests

🔒 Security Tip: In multi-Agent environments, keep team_mode: false unless you need team ranking.

Future Roadmap

Memory → Growth → Evolution → Alive

• v6 Series (Current) — 5-layer memory + Knowledge Palace + Intimacy + Native memory import
• v7 Evolution (Planned) — Agent self-creates skills, from executor to creator
• Long-term — Continuous cognitive architecture evolution

GitHub: https://github.com/anima-aios/anima | Apache 2.0

---

✨ v6.2.4 New Features (Current Version)

🤝 self-improving-agent Compatibility

Silent Detection:

• Automatically scans .learnings/ directory if exists
• No prompts if user hasn't installed self-improving
• Extracts high-value learning records to L2 facts
• Rewards EXP for learning behavior

Compatibility:

• Users with self-improving: Auto-sync enabled
• Users without: No impact, normal operation

🏆 Team Ranking Built-in

Features:

• Auto-scans all agents' cognitive profiles
• Generates rankings by EXP/Level/5-Dimensions
• Outputs Markdown + JSON formats
• Scheduled daily at 00:00

Ranking Types:

• EXP Ranking (Top 10)
• Level Ranking (Top 10)
• Cognitive Score Ranking (Top 10)
• 5-Dimension Rankings (Each dimension Top 10)

Output:

• /home/画像/shared/团队排行榜_{date}.md
• /home/画像/shared/团队排行榜_{date}.json

---

✨ v6.2.3 New Features (Previous Version)

🔒 Security & Privacy Fixes

• Version Unification - init.py updated from 6.1.2 to 6.2.1
• Privacy Default Protection - team_mode changed to false, no scanning of other Agents' data
• Documentation Transparency - Changed "zero-intrusion" to "low-intrusion", clarified background behavior
• New Privacy Section - Added background behavior section and config privacy tips
• Install Prompt Optimization - post-install.sh adds sensitive feature disable guide

---

✨ v6.2.0 New Features

🏗️ 5-Layer Memory Architecture

• L1 Working Memory: Auto-listens to OpenClaw memory/ directory changes, zero-intrusion sync
• L2 Episodic Memory: Event archiving, LLM quality assessment (S/A/B/C)
• L3 Semantic Memory: LLM-driven knowledge distillation + semantic deduplication
• L4 Knowledge Palace: Spatial knowledge organization + Pyramid distillation (Instance→Rule→Pattern→Ontology)
• L5 Metacognition: Memory decay function + Health system + 5-D profile

🔌 Native Integration with OpenClaw

• memory_watcher: Based on watchdog library, auto-detects inotify/FSEvents/WinAPI
• Agent's daily memory writes automatically trigger Anima sync, completely imperceptible
• Solves FB-008: Memory sync breakage issue

🏛️ Knowledge Palace

• Palace → Floor → Room → Location → Item, 5-level spatial structure
• Default 4 knowledge rooms + _inbox fallback
• LLM intelligent classification + delayed debounce scheduler (organize after typing stops)

🔺 Pyramid Knowledge Organization

• Instance → Rule → Pattern → Ontology, 4-layer bottom-up distillation
• Trigger Condition: Auto-distills when ≥3 instances of same topic
• Advanced: Distills to Pattern when ≥5 rules of same topic
• Conservative mode: auto_distill=false by default, controlled by config switch

📉 Memory Decay Function

• Based on Ebbinghaus forgetting curve + AI scenario adaptation
• Review = Access: Automatically refreshes on each memory_search hit
• Review recommendations + Forgetting alerts + Archive markers

🏥 Health System (5 Modules)

• manager: Master scheduler, Doctor command entry point
• hygiene: Data integrity checks + deduplication + cleanup
• correction: Auto-detects and fixes common data issues
• evolution: Daily auto-distillation in early morning (L2→L3 + Palace classification + Pyramid distillation)
• abstraction: Cross-room knowledge association discovery

🤖 LLM Intelligent Processing

• Quality assessment / Deduplication analysis / Palace classification all support LLM
• Multi-model config: Uses current Agent model by default (most accurate), configurable per task
• Auto-degrades to rule mode when LLM unavailable

---

✨ Retained Features (v5)

🧠 Enhanced Memory Management

• Multi-layer Sync: OpenClaw Memory + Anima Facts + EXP History
• Intimacy Rewards: Auto-gains intimacy when writing memory
• Intelligent Deduplication: Automatically avoids duplicate records

📊 5-Dimensional Cognitive Profile

• Internalization: Knowledge absorption and understanding ability
• Application: Knowledge transfer and practical ability
• Creation: Knowledge integration and innovation ability
• Metacognition: Self-reflection and monitoring ability
• Collaboration: Teamwork and mutual assistance ability

🎮 Gamified Growth

• Level System: From Lv.1 Novice to Lv.100 Lifetime Achievement
• Daily Quests: 3 challenges per day, extra intimacy on completion
• Progress Tracking: Visual upgrade progress bar

🏆 Team Leaderboard

• Intimacy Ranking: Based on fair normalized algorithm
• Real-time Competition: Track ranking changes and gaps

---

🛠️ Architecture

Agent Daily Work (OpenClaw write/edit/memory_write)
       │
       ▼  watchdog listens, zero-intrusion
 L1 Working Memory ── workspace/memory/*.md
       │沉淀
       ▼
 L2 Episodic Memory ── facts/episodic/ (LLM quality assessment)
       │提炼
       ▼
 L3 Semantic Memory ── facts/semantic/ (LLM dedup + association)
       │结构化
       ▼
 L4 Knowledge Palace ── palace/rooms/ (LLM classification + debounce)
    Pyramid   ── pyramid/ (Instance→Rule→Pattern→Ontology)
       │反思
       ▼
 L5 Metacognition ── 5-D Profile + Intimacy + Decay + Health

---

📁 Module List

core/ (Core Modules)

Module	Version	Description
memory_watcher.py	v6.0	OpenClaw memory file monitoring + auto-sync
fact_store.py	v6.0	L2/L3 unified fact storage layer
distill_engine.py	v6.0	L2→L3 LLM-driven distillation engine
palace_index.py	v6.0	Memory Palace spatial index
pyramid_engine.py	v6.0	Pyramid knowledge organization engine
palace_classifier.py	v6.0	Palace classification scheduler (debounce)
decay_function.py	v6.0	Ebbinghaus memory decay calculation
cognitive_profile.py	v5→v6	5-D cognitive profile generator
exp_tracker.py	v5	Intimacy tracking
level_system.py	v5	Level system
daily_quest.py	v5	Daily quests
memory_sync.py	v5→v6	Memory sync (path hardcoding fixed)

health/ (Health System)

Module	Version	Description
manager	v6.0	Master scheduler + Doctor entry
hygiene	v6.0	Data hygiene (integrity + dedup + cleanup)
correction	v6.0	Auto-correction
evolution	v6.0	Daily evolution (early morning auto-distillation)
abstraction	v6.0	Knowledge abstraction (cross-room association)

---

⚙️ Configuration

Config file path: ~/.anima/config/anima_config.json

{
  "facts_base": "/home/画像",
  "agent_name": "auto",
  "llm": {
    "provider": "current_agent",
    "models": {
      "quality_assess": "current_agent",
      "dedup_analyze": "current_agent",
      "palace_classify": "current_agent"
    }
  },
  "palace": {
    "classify_mode": "deferred",
    "poll_interval_minutes": 30,
    "quiet_threshold_seconds": 60,
    "retry_delay_seconds": 60
  },
  "pyramid": {
    "auto_distill": false,
    "distill_threshold": 3
  },
  "team_mode": false
}

Key Configuration:

Config	Description	Default	Recommendation
team_mode	Scan other Agents' data for team ranking	false	Keep disabled in multi-Agent env
facts_base	Fact data storage path	/home/画像	Can customize to private directory
agent_name	Agent name	Auto-detect	Usually no modification needed

🔐 Privacy Tip: With team_mode: false, Anima only processes current Agent's data, won't access other Agents' files.

---

🧪 Testing

# Install dependencies (required for memory_watcher)
pip install "watchdog>=3.0.0"

# Run integration tests (37 checks)
python3 tests/test_integration_v6.py

---

The architecture can only evolve, not degenerate. — Liu Wen's Iron Rule First be honest, then iterate. Code must match the hype. — Qing He

---

---

Anima-AIOS v6.0 (中文版)

让成长可见，让认知可量 | Making Growth Visible, Making Cognition Measurable

为你的 AI Agent 添加五层记忆架构、知识宫殿、认知成长和自动进化能力。

---

描述

你的 Agent 每天都在「重新活一次」。Anima 改变这一点。

Anima（拉丁语「灵魂」）为 OpenClaw Agent 提供五层记忆架构，模拟人类认知发展过程，让 Agent 能记住经历、沉淀知识、形成认知、持续成长。

核心能力

• 🧠 五层记忆架构 L1→L5 — 工作记忆→情景→语义→知识宫殿→元认知
• 🏛️ 知识宫殿 — 5 级空间结构 + LLM 智能分类，市面独有
• 🔺 金字塔知识组织 — 实例→规则→模式→本体，4 层自动提炼
• 📉 Ebbinghaus 记忆衰减 — 科学遗忘曲线 + 智能复习推荐
• 👁️ 低侵入 Watchdog — 可选自动记忆监听，无需修改 Agent 代码
• 🧬 五维认知画像 — 内化力 · 应用力 · 创造力 · 元认知 · 协作力
• 🏥 健康系统 — 5 大模块保障数据可靠性
• 🔄 v6.2 原生记忆导入 — 一键导入 OpenClaw 记忆，解决冷启动

安装

clawhub install anima-aios
pip install watchdog  # 可选：启用自动记忆监听

低侵入配置，可选后台监听，安装后建议运行自检。

💡 提示：支持 LLM 模式（智能分类/去重/质量评估），无 LLM 时自动降级为规则模式。

⚠️ 后台行为与隐私说明

后台功能（默认关闭或可选）：

功能	说明	默认状态	关闭方法
memory_watcher	基于 watchdog 的文件系统监听，自动同步记忆	需手动启用	不安装 watchdog 或在配置中禁用
每日进化	凌晨自动提炼 L2→L3 记忆	需配置 cron	不配置 cron 任务
团队排行	扫描其他 Agent 的认知画像	❌ 默认关闭	team_mode: false（默认已关闭）

隐私保护：

• team_mode 默认为 false，不会扫描其他 Agent 数据
• 如需启用团队排行，请在配置中手动设置 team_mode: true
• 所有数据处理均在本地完成，无网络请求

🔒 安全提示：多 Agent 环境下，建议保持 team_mode: false，除非你需要团队排行功能。

未来蓝图

记忆 → 成长 → 进化 → 活着

• v6 系列（当前） — 五层记忆 + 知识宫殿 + 亲密度 + 原生记忆导入
• v7 进化（规划中） — Agent 自创技能，从执行者变创造者
• 远期 — 认知架构持续演进

GitHub: https://github.com/anima-aios/anima | Apache 2.0

---

✨ v6.2.3 新增功能（当前版本）

🔒 文档透明度提升

多平台路径说明：

• Linux: /home/画像（多 Agent 共享）
• macOS: ~/画像
• Windows: ~/画像
• 环境变量：ANIMA_FACTS_BASE 可覆盖

网络调用透明说明：

• LLM API 调用（可选，用户可控）
• 支持本地部署（无网络）
• 默认降级为规则模式

脚本用途说明：

• post-install.sh - 安装时复制 Core
• refresh-quests.sh - 刷新每日任务
• sync-memory.sh - 定时同步记忆
• show-progress.sh - 显示认知进度
• 全部本地操作，无网络调用

环境变量统一：

• 统一为 ANIMA_* 前缀
• OPENCLAW_WORKSPACE 兼容（deprecated 警告）

🔧 环境变量统一

变更前：

• ANIMA_FACTS_BASE ✅
• ANIMA_AGENT_NAME ✅
• OPENCLAW_WORKSPACE ⚠️
• WORKSPACE ❌

变更后：

• ANIMA_FACTS_BASE ✅ 主要
• ANIMA_AGENT_NAME ✅ 主要
• OPENCLAW_WORKSPACE ⚠️ 兼容（deprecated 警告）

---

✨ v6.2.2 新增功能（上一版本）

🔧 per-Agent 配置覆盖

问题： 多 Agent 场景下，全局配置无法满足个性化需求（如不同的 LLM 配置、五维权重）

解决方案： 支持 per-Agent 配置覆盖

配置结构：

~/.anima/config/
├── config.json          # 全局默认配置（所有 Agent 共享）
└── agents/
    ├── Z.json           # Z 的覆盖配置（只写差异）
    ├── 方秋.json        # 方秋的覆盖配置
    └── ...

配置合并逻辑：

最终配置 = 代码默认值 + 全局配置 + Agent 覆盖配置

示例：

全局配置 (config.json):

{
  "facts_base": "/home/画像",
  "llm": { "provider": "current_agent" },
  "weights": { "creation": 0.25 }
}

Z 的覆盖配置 (agents/Z.json):

{
  "llm": { "provider": "bailian", "models": { "quality_assess": "qwen-max" } },
  "weights": { "creation": 0.30 }
}

最终 Z 的配置 = 全局 + Z 覆盖（深度合并）

移除： "agent" 字段（改为运行时自动检测）

优先级：

1. 环境变量（最高）
2. Agent 覆盖配置
3. 全局配置
4. 代码默认值

---

✨ v6.2.1 新增功能（上一版本）

🔒 安全与隐私修复

• 版本号统一 - init.py 从 6.1.2 更新为 6.2.1
• 隐私默认保护 - team_mode 默认改为 false，不扫描其他 Agent 数据
• 文档透明度提升 - 修改"零侵入"为"低侵入"，明确说明后台行为
• 新增隐私说明 - 添加后台行为说明章节和配置隐私提示
• 安装提示优化 - post-install.sh 添加敏感功能关闭指南

---

✨ v6.2.0 新增功能

🏗️ 五层记忆架构

• L1 工作记忆：自动监听 OpenClaw memory/ 目录变化，零侵入同步
• L2 情景记忆：事件归档，LLM 质量评估（S/A/B/C）
• L3 语义记忆：LLM 驱动的知识提炼 + 语义去重
• L4 知识宫殿：空间化知识组织 + 金字塔知识提炼（实例→规则→模式→本体）
• L5 元认知层：记忆衰减函数 + 健康系统 + 五维画像

🔌 与 OpenClaw 原生打通

• memory_watcher：基于 watchdog 库，自动识别 inotify/FSEvents/WinAPI
• Agent 日常写 memory 自动触发 Anima 同步，完全无感知
• 解决 FB-008：记忆同步断裂问题

🏛️ 知识宫殿（Knowledge Palace）

• 宫殿 → 楼层 → 房间 → 位置 → 物品，五级空间结构
• 默认 4 个知识房间 + _inbox 兜底
• LLM 智能分类 + 延迟防抖调度器（等笔停了再整理）

🔺 金字塔知识组织

• 实例 → 规则 → 模式 → 本体，四层自底向上提炼
• 触发条件： 同一主题 ≥ 3 条实例时自动触发规则提炼
• 进阶提炼： 同一主题 ≥ 5 条规则时触发模式提炼
• 保守模式：默认 auto_distill=false，config 开关控制

📉 记忆衰减函数

• 基于 Ebbinghaus 遗忘曲线 + AI 场景适配
• 复习 = 访问：每次 memory_search 命中自动刷新
• 复习推荐 + 即将遗忘提醒 + 可归档标记

🏥 健康系统（5 个模块）

• manager：总调度，Doctor 命令入口
• hygiene：数据完整性检查 + 去重 + 清理
• correction：自动检测并修复常见数据问题
• evolution：每日凌晨自动提炼（L2→L3 + 宫殿分类 + 金字塔提炼）
• abstraction：跨房间知识关联发现

🤖 LLM 智能处理

• 质量评估 / 去重分析 / 宫殿分类均支持 LLM
• 多模型配置：默认用当前 Agent 模型（最准），可按任务配置不同模型
• LLM 不可用时自动降级为规则模式

---

✨ 保留功能（v5）

🧠 增强记忆管理

• 多层同步：OpenClaw Memory + Anima Facts + EXP History
• 亲密度奖励：写记忆自动获得亲密度
• 智能去重：自动避免重复记录

📊 五维认知画像

• 内化力：知识吸收和理解能力
• 应用力：知识迁移和实践能力
• 创造力：知识整合和创新能力
• 元认知：自我反思和监控能力
• 协作力：团队合作和互助能力

🎮 游戏化成长

• 等级系统：从 Lv.1 新手到 Lv.100 终身成就
• 每日任务：每天 3 个挑战，完成获得额外亲密度
• 进度追踪：可视化升级进度条

🏆 团队排行榜

• 亲密度排行：基于公平归一化算法排名
• 实时竞争：追踪排名变化和差距

---

🛠️ 架构

Agent 日常工作（OpenClaw write/edit/memory_write）
       │
       ▼  watchdog 监听，零侵入
 L1 工作记忆 ── workspace/memory/*.md
       │ 沉淀
       ▼
 L2 情景记忆 ── facts/episodic/（LLM 质量评估）
       │ 提炼
       ▼
 L3 语义记忆 ── facts/semantic/（LLM 去重 + 关联）
       │ 结构化
       ▼
 L4 知识宫殿 ── palace/rooms/（LLM 分类 + 延迟防抖）
    金字塔   ── pyramid/（实例→规则→模式→本体）
       │ 反思
       ▼
 L5 元认知层 ── 五维画像 + 亲密度 + 衰减 + 健康

---

📁 模块清单

core/（核心模块）

模块	版本	说明
memory_watcher.py	v6.0	OpenClaw 记忆文件监听 + 自动同步
fact_store.py	v6.0	L2/L3 统一事实存储层
distill_engine.py	v6.0	L2→L3 LLM 驱动提炼引擎
palace_index.py	v6.0	记忆宫殿空间索引
pyramid_engine.py	v6.0	金字塔知识组织引擎
palace_classifier.py	v6.0	宫殿分类调度器（延迟防抖）
decay_function.py	v6.0	Ebbinghaus 记忆衰减计算
cognitive_profile.py	v5→v6	五维认知画像生成器
exp_tracker.py	v5	亲密度追踪
level_system.py	v5	等级系统
daily_quest.py	v5	每日任务
memory_sync.py	v5→v6	记忆同步（已修复路径硬编码）

health/（健康系统）

模块	版本	说明
manager	v6.0	总调度 + Doctor 入口
hygiene	v6.0	数据卫生（完整性 + 去重 + 清理）
correction	v6.0	自动纠错
evolution	v6.0	每日进化（凌晨自动提炼）
abstraction	v6.0	知识抽象（跨房间关联）

---

⚙️ 配置 (v6.2.2)

配置结构

全局配置 (~/.anima/config/config.json):

{
  "version": "6.2.2",
  "facts_base": "/home/画像",
  "llm": {
    "provider": "current_agent",
    "models": {
      "quality_assess": "current_agent",
      "dedup_analyze": "current_agent",
      "palace_classify": "current_agent"
    }
  },
  "palace": {
    "classify_mode": "deferred",
    "poll_interval_minutes": 30,
    "quiet_threshold_seconds": 60,
    "retry_delay_seconds": 60
  },
  "pyramid": {
    "auto_distill": false,
    "distill_threshold": 3
  },
  "team_mode": false
}

Agent 覆盖配置 (~/.anima/config/agents/{agent_name}.json):

{
  "_comment": "只写与全局配置的差异",
  "llm": {
    "provider": "bailian",
    "models": {
      "quality_assess": "qwen-max"
    }
  },
  "weights": {
    "creation": 0.30
  }
}

配置优先级

优先级	来源	说明
1	环境变量	ANIMA_FACTS_BASE, ANIMA_TEAM_MODE 等
2	Agent 覆盖配置	~/.anima/config/agents/{agent_name}.json
3	全局配置	~/.anima/config/config.json
4	代码默认值	config_loader.py 中的 DEFAULT_CONFIG

关键配置说明

配置项	说明	默认值	建议
team_mode	是否扫描其他 Agent 数据生成团队排行	false	多 Agent 环境保持关闭
facts_base	事实数据存储路径	/home/画像	可自定义到私有目录
llm.provider	LLM 提供商	current_agent	可用 bailian, openai 等
pyramid.auto_distill	是否启用金字塔自动提炼	false	数据量大时可启用

🔐 隐私提示：team_mode: false 时，Anima 仅处理当前 Agent 的数据，不会访问其他 Agent 文件。

💡 提示：Agent 名称自动检测（环境变量 → OpenClaw 上下文 → SOUL.md → 兜底），无需手动配置。

---

🧪 测试

# 安装依赖（memory_watcher 需要）
pip install "watchdog>=3.0.0"

# 运行集成测试（37 项检查）
python3 tests/test_integration_v6.py

---

架构只能演进，不能退化。—— 立文铁律 先诚实，再迭代。代码要配得上宣传。—— 清禾