Human Like Memory

模仿人类记忆机制的智能上下文管理系统，支持向量化检索、智能摘要、三层存储（HOT/WARM/COLD）、按需注入

Security Scan

The skill's files, instructions, and dependencies are coherent with a local, Node.js-based memory/embedding system; it installs npm packages and downloads a ~200MB model but does not request unrelated credentials or unusual system access.

Purpose & Capability

Name/description (vector search, compression, injection) match the included code (vector-engine.js, compression-engine.js, injection-engine.js), package.json dependency (@xenova/transformers), and config.json settings. Files and docs implement the advertised features.

Instruction Scope

SKILL.md and ancillary docs instruct automatic dependency installation and running test scripts (npm install @xenova/transformers; node test-vector-engine.js). Instructions reference downloading a model from Hugging Face and optional local cache settings. All referenced actions are within the skill's stated purpose, but the docs assume the platform will run background installs and model downloads.

Install Mechanism

There is no platform install spec, but an install.sh and package.json/package-lock indicate npm-based installs. Installing @xenova/transformers (npm) and downloading the Xenova/bge-m3 model are expected for local embeddings; this is a moderate supply-chain risk (npm packages and model download) but uses standard registries (npm, Hugging Face) rather than obscure hosts.

Credentials

The skill declares no required environment variables or credentials. Optional env vars in docs (TRANSFORMERS_CACHE, HF_ENDPOINT, NODE_OPTIONS) are reasonable for model caching and performance tuning and are not secrets. No requests for cloud keys or unrelated credentials are present.

Persistence & Privilege

Skill is user-invocable and not forced-always. Installation and model cache are confined to its skill directory and usual caches (~/.cache). It does not modify other skills or request system-wide config changes.

Assessment

This skill appears internally consistent with a local-memory/embedding tool, but note these practical risks before installing: - Dependency and model downloads: install.sh / SKILL.md will trigger npm install (fetching @xenova/transformers and transitive packages) and download a ~200MB model from Hugging Face. That is expected for embeddings but is a supply-chain/network activity — review package-lock.json and the install script if you need to vet dependencies. - Unknown publisher / no homepage: the registry metadata has no homepage and an unfamiliar owner ID. If you require higher assurance, review the code files (they are included) or run installation in an isolated environment/container before enabling on production agents. - Resource/network considerations: first run requires disk (~200–300MB) and network access. If you must avoid outbound downloads, set useVectorSearch=false in config.json or run in an environment with an approved model mirror. - Recommended precautions: inspect install.sh and package-lock.json, run npm install and test scripts in a sandbox, and verify that the behavior (model downloads, cache path) matches your policies. If you are comfortable with npm/Hugging Face dependencies and can sandbox the install, the skill is consistent with its stated purpose.

Human-Like Memory Skill v1.1.1 🧠

下一代智能记忆管理系统 — 向量化检索 + 智能压缩 + 按需注入

✨ v1.2.0 新特性

功能	说明	Token 节省
🧠 向量化检索	语义搜索，相似度>0.7	-40%
🗜️ 智能摘要	3 级压缩（原文/摘要/关键词）	-20%
💉 按需注入	动态加载 Top 5 相关记忆	-30%
📦 本地运行	无需 API，零配置	100% 隐私
⚡ 自动安装	依赖自动安装，无需手动操作	100% 便捷
🚀 性能优化	缓存 + 并行处理，提速 50%	-50% 延迟

总体效果：记忆容量提升 3 倍，Token 占用减少 60%

---

🚀 零配置使用

重要改进: v1.1.1 起，技能支持自动依赖安装！

安装流程

# 用户只需执行这一条命令
clawhub install human-like-memory-cn

# OpenClaw 会自动:
# 1. 下载技能文件
# 2. 检测依赖 (@xenova/transformers)
# 3. 后台执行 npm install
# 4. 完成后技能可用

无需手动执行 npm install！

---

核心功能

1. 向量化记忆检索 — 语义搜索，相似度>0.7 视为相关
2. 智能摘要压缩 — 按重要性自动压缩（原文/摘要/关键词）
3. 按需上下文注入 — 每轮对话动态加载最相关记忆
4. 三层记忆架构 — HOT（7 天）/ WARM（7-30 天）/ COLD（永久）
5. 记忆遗忘机制 — 自动清理低价值信息