Install
openclaw skills install anima-aiosAn AI Agent cognitive growth system built on the native OpenClaw architecture. It provides agents with persistent memory management, visual intimacy progress...
openclaw skills install anima-aios🌐 Language / 语言切换:
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.
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.
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 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' datateam_mode: true in config🔒 Security Tip: In multi-Agent environments, keep
team_mode: falseunless you need team ranking.
Memory → Growth → Evolution → Alive
GitHub: https://github.com/anima-aios/anima | Apache 2.0
Silent Detection:
.learnings/ directory if existsCompatibility:
Features:
Ranking Types:
Output:
/home/画像/shared/团队排行榜_{date}.md/home/画像/shared/团队排行榜_{date}.jsonAgent 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 | 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) |
| 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) |
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.
# 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
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。
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。
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team_mode: true🔒 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。
记忆 → 成长 → 进化 → 活着
GitHub: https://github.com/anima-aios/anima | Apache 2.0
多平台路径说明:
/home/画像(多 Agent 共享)~/画像~/画像ANIMA_FACTS_BASE 可覆盖网络调用透明说明:
脚本用途说明:
环境变量统一:
ANIMA_* 前缀OPENCLAW_WORKSPACE 兼容(deprecated 警告)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)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:
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 | 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) |
| 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) |
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
}
}
| 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 |
| 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。
# 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