Mem0 Memory

mem0 本地记忆层完整实现（增强版）。语义记忆存储/检索/管理，WAL 协议，SESSION-STATE，多级记忆（User/Session/Agent）。参考 ZejunCao/bilibili_code Mem0框架解读优化。

Security Scan

The skill's documentation describes a local-memory system that reads/writes persistent files and depends on local services (Ollama, MiniMax, Chroma), but the skill package declares no binaries, no config paths, and provides no code—these mismatches and the privacy-sensitive storage behavior are concerning.

Purpose & Capability

The SKILL.md promises a full local memory stack (Python wrapper mem0_wrapper.py, Ollama embedder, MiniMax LLM, Chroma vector DB) and points at a concrete filesystem path (D:\autoclaw\结果\mem0\). However, the skill metadata declares no required binaries, no config paths, and includes no code or install steps. A legitimate implementation would require at minimum: Python/runtime, the mem0_wrapper.py script and supporting files, and locally running services (Ollama, Chroma/DB files). The absence of these requirements in the package metadata is an incoherence.

Instruction Scope

The instructions direct the agent to run Python commands against files in a specific local path, scan every incoming message for multiple data types, and persist user/session/agent memories to disk (including URLs and file paths). That scope involves reading/writing local files and continuously processing user messages for storage; these actions are broader and more privacy-sensitive than the declared skill footprint. There is no guidance in the metadata about what files will be created, retention, encryption, or where SESSION-STATE.md / WORKING-BUFFER.md live, so the runtime instructions exceed what the package declares.

Install Mechanism

The skill is instruction-only (no install spec), which is low risk from an installer perspective. However, it depends on several external components (Ollama running as a service, local LLM models, Chroma DB, and a Python wrapper) that are not installed or provided by the skill. That mismatch means the agent or operator must have preinstalled components — this should be declared and verified before use.

Credentials

No environment variables or credentials are requested, which superficially looks safe. But the SKILL.md expects access to a specific local filesystem location and to persist potentially sensitive user data (preferences, experiences, URLs, file paths). The lack of declared config paths and permission expectations in the metadata is disproportionate: the skill requires filesystem and local-service access but doesn't declare it. Storing personal data persistently without documented controls (encryption, retention policy, access controls) is a privacy risk.

Persistence & Privilege

The skill persists memories to disk and relies on SESSION-STATE.md and WORKING-BUFFER.md for state recovery; it does not set always:true and does not modify other skills. Persisting user memory is expected for this purpose, but the skill's metadata fails to document where/how data is persisted and protected. The per-message WAL scanning behavior (scan every message for triggers) increases privacy sensitivity and should be explicitly consented and controlled.

What to consider before installing

Key points to consider before installing or using this skill: - The package contains only a SKILL.md; it does not include the Python scripts or binaries it references. Verify that the referenced files (e.g., D:\autoclaw\结果\mem0\mem0_wrapper.py) and services (Ollama, the MiniMax model, Chroma DB) actually exist and come from a trusted source before running anything. - This skill will scan every message and persist memories (including URLs, file paths, names, preferences) to local files. That can capture sensitive personal or system information — confirm retention, access controls, and whether data is encrypted at rest. - The metadata does not declare required runtimes (Python) or config paths. Ask the author for a manifest: exact runtime requirements, file layout, where data is stored, and a copy of mem0_wrapper.py to review. - Do not point this skill at directories with sensitive files (password stores, SSH keys, private documents). Consider running it in an isolated environment or VM and review/read the actual code before giving it access to your real user data. - If you proceed, require explicit confirmation steps for destructive commands (reset/delete_all) and log all actions. If you cannot obtain the implementation or a trustworthy origin, avoid installing the skill.

mem0 Memory 🧠 — 完整实现（增强版）

组件

组件	技术	作用
LLM	MiniMax-M2.7	记忆生成与提取
Embedder	Ollama nomic-embed-text	本地语义向量化（完全离线）
向量库	Chroma	本地向量存储
Ollama	开机自启	向量服务

文件路径：D:\autoclaw\结果\mem0\

增强功能（参考 bilibili_code 课件）

chat_with_memories 对话模式

mem0_wrapper.py chat 命令实现了完整的记忆增强对话流程：

1. search 检索相关记忆
2. 将记忆注入 system prompt
3. 调用 MiniMax LLM 生成回复

中文记忆提取

内置 FACT_RETRIEVAL_PROMPT，自动将用户对话提取为中文记忆条目。

多级记忆支持

--run 参数支持 Session 级别记忆，--agent 支持 Agent 级别记忆。

完整 API（10 个命令）

# 增删改查
python D:\autoclaw\结果\mem0\mem0_wrapper.py add <user_id> "<内容>"              # 添加记忆（自动提取）
python D:\autoclaw\结果\mem0\mem0_wrapper.py search <user_id> "<query>"          # 语义搜索
python D:\autoclaw\结果\mem0\mem0_wrapper.py get_all <user_id>                 # 全部记忆
python D:\autoclaw\结果\mem0\mem0_wrapper.py get <memory_id>                    # 单条记忆
python D:\autoclaw\结果\mem0\mem0_wrapper.py update <memory_id> "<新内容>"     # 更新记忆
python D:\autoclaw\结果\mem0\mem0_wrapper.py delete <memory_id>                # 删除记忆
python D:\autoclaw\结果\mem0\mem0_wrapper.py history <memory_id>               # 修改历史

# 记忆增强对话（新增）
python D:\autoclaw\结果\mem0\mem0_wrapper.py chat <user_id> "<问题>"            # 检索+注入+LLM回答

# 重置
python D:\autoclaw\结果\mem0\mem0_wrapper.py delete_all <user_id>               # 清空用户记忆
python D:\autoclaw\结果\mem0\mem0_wrapper.py reset                              # 清空全部记忆（慎用）

# 可选参数
--limit N    # 限制返回数量（search/chat 默认5）
--run <id>  # 指定会话ID（session级别记忆）
--agent <id> # 指定智能体ID（agent级别记忆）

多级记忆架构

mem0 存储层
├── User Memory（user_id=main_user）  ← 跨会话长期记忆
├── Session Memory（run_id）           ← 单会话临时记忆
└── Agent Memory（agent_id）           ← 多智能体共享记忆

当前实现：User Memory（main_user），SESSION-STATE.md 作为 Session 层补充。

WAL 触发扫描（每消息必做）

触发类型	示例	存储
偏好	"我喜欢..." / "不要..."	mem0 add
经历	"我去了..." / "我做过..."	mem0 add
重要事实	"我有..." / "我是..."	mem0 add
纠正	"不是Y，是X" / "其实..."	SESSION-STATE
决定	"就用X" / "去..."	SESSION-STATE + mem0
数字/日期	具体数字+单位、日期	SESSION-STATE
URL/路径	链接、文件路径	SESSION-STATE
专有名词	名字、地点、公司、产品	判断后存储

优先级：SESSION-STATE > mem0

完整工作流

收到消息时

扫描类型
  ├─ [偏好/经历/事实] → mem0 add
  ├─ [纠正/决定/数字/URL] → SESSION-STATE.md
  └─ [闲聊/无价值] → 不存
回复用户

回复后

上下文使用率 > 60%？
  └─ 是 → WORKING-BUFFER.md 激活

截断恢复（下次会话）

1. mem0 get_all → 恢复长期语义记忆
2. mem0 chat → 主动询问是否继续上次任务
3. SESSION-STATE.md → 恢复当前任务状态
4. WORKING-BUFFER.md → 恢复危险区对话

参考资料

• ZejunCao/bilibili_code（Mem0框架解读）：https://github.com/ZejunCao/bilibili_code
  • mcp_server.py — FastMCP + mem0 REST 服务架构参考
  • prompts_zh.py — 中文提示词模板（USER_MEMORY_EXTRACTION_PROMPT 等）
  • demo.ipynb — chat_with_memories 完整流程
  • 踩坑记录.md — OpenMemory 调试踩坑

注意事项

• Ollama 必须后台运行（已设开机自启）
• 添加需要 LLM（慢），搜索只需 embedder（快）
• SESSION-STATE 优先于 mem0
• 危险操作（reset/delete_all）需要二次确认
• memory_id 是记忆的唯一标识，用于 update/delete/history