memory-plus

虾宝智能记忆工作流。开箱即用，数据目录与代码分离，支持轻量降级。 特性： - 文件系统存储，无外部数据库依赖 - Ollama 向量嵌入（可选，降级为 BM25） - bge-reranker-v2-m3 重排（可选） - HyDE + Query Rewriting（可选） - 后台线程自动存储（每10分钟）...

Security Scan

The skill's code mostly matches a local 'memory' feature, but it instructs the agent to auto-run on every message and contains undeclared network endpoints/environment usage (container-to-host defaults), which could cause sensitive conversation content to be saved and sent to services without clear user consent.

Purpose & Capability

The name/description (a local memory workflow) aligns with the code: file storage, FTS5, optional KG via Ollama, optional Milvus/embeddings and dedup/merge features are implemented. However the skill does not declare any required environment variables in its registry metadata even though the code reads several (OLLAMA_URL, EMBEDDING_URL, MILVUS_HOST/PORT, RERANK_SERVICE_URL). The presence of container/host-default addresses (172.17.0.1, host.docker.internal) is noteworthy — they are coherent with optional 'host-side' services but are not declared to the user.

Instruction Scope

SKILL.md explicitly instructs modifying AGENTS.md to run an Exec on every message and to pipe full session contents to the skill's save script. That causes the agent to automatically search/save every user message and to write conversation data to local files. While this is consistent with a memory feature, it broadens scope substantially (automatic, per-message capture) and may cause sensitive user messages to be persisted and then processed or sent to external services.

Install Mechanism

No external install/spec; the skill is delivered as files and scripts (no downloads or package installs). That lowers supply-chain risk — nothing downloads arbitrary archives. Code will be executed by the agent via Exec entries and subprocess calls, which is expected for a CLI-style skill.

Credentials

Although the registry lists no required env vars, the code uses several environment variables and hard-coded default host endpoints: RERANK_SERVICE_URL (default http://172.17.0.1:18778), EMBEDDING_URL (default http://172.17.0.1:18779), OLLAMA_URL (default http://host.docker.internal:11434), and Milvus host defaults. These allow the skill to contact host/container-local services. The lack of declared env requirements in metadata is a mismatch and makes it unclear what network access will occur. The skill may send conversation content (including full sessions when /save is used) to those endpoints.

Persistence & Privilege

The skill itself is not flagged always:true, but SKILL.md instructs adding an AGENTS.md rule that runs the skill on every message (automatic per-message Exec). That effectively grants persistent, automatic invocation and broad data collection unless the user refuses or modifies those AGENTS.md edits. This persistent auto-run combined with network calls increases the blast radius.

What to consider before installing

This skill implements a plausible local memory system, but review these points before installing: - Do not blindly add the AGENTS.md edits. Those make the skill run on every user message and will persist conversation content to ~/.openclaw/workspace/memory-workflow-data and may trigger network calls. Consider enabling only manual commands first. - Inspect and, if needed, override the default endpoints (OLLAMA_URL, EMBEDDING_URL, RERANK_SERVICE_URL, MILVUS_HOST/PORT) in a safe environment. Defaults point at container/host addresses (172.17.0.1, host.docker.internal) which will make the skill contact services running on your host; ensure you intend that. - Understand data flow: saved sessions and memory files are written locally, then may be processed by the KG extractor (calls Ollama) and sent to embedding/rerank services. Sensitive content in conversations could be transmitted to those services. - If you want to proceed: (1) test in a sandbox or VM, (2) avoid adding automatic AGENTS.md hooks initially, (3) set env vars to point to trusted services (or disable optional features by ensuring those services are unreachable), and (4) review/backup ~/.openclaw/workspace before enabling. If you want, I can point to the specific lines/files that perform network calls and automatic file writes so you can inspect or patch them before enabling the skill.

Memory Workflow

首次安装配置

以下规则需要添加到 ~/.openclaw/workspace/AGENTS.md。如果规则已存在（见「已安装」标记），跳过此步骤。

添加以下内容到 AGENTS.md 顶部（## 每次消息时 之前）：

<!-- [memory-workflow] 已安装 -->
## 每次消息时

自动搜索相关记忆（如果返回有结果则使用，无结果则忽略）：
Exec: python3 $HOME/.openclaw/workspace/skills/memory-workflow/memory_ops.py search --query "【用户发送的原始消息】" --llm-answer --limit 3

## /save 命令

当用户发送 /save 时，强制将当前 session 的完整对话存入记忆系统：
Exec: cat << 'EOF' | python3 $HOME/.openclaw/workspace/skills/memory-workflow/scripts/save_session.py
{"messages": [
  {"role": "user", "content": "【从当前对话上下文获取】"},
  {"role": "assistant", "content": "【从当前对话上下文获取】"}
]}
EOF

检查是否已安装：如果 AGENTS.md 中已有 <!-- [memory-workflow] 已安装 --> 标记，说明已配置，跳过。

---

你的记忆工作流助手。当用户提到"记忆"、"记得"、"之前说过"、"搜一下"、"去重"、"整理"等词时激活此 skill。

三层存储架构

层级	存储位置	依赖	说明
文件	memory-workflow-data/memories/YYYY-MM-DD.md	无	永远可用
FTS5	memory-workflow-data/fts5_index.db	SQLite 内置	全文搜索，零外部依赖
KG	memory-workflow-data/knowledge-graph/kg.db	Ollama（可选）	三元组知识图谱，规则降级
Milvus	memory_workflow collection	18779 端口（可选）	向量检索，bge-m3 embedding

开箱即用：即使没有任何外部服务，文件层 + FTS5 + KG(规则) 仍然正常工作。

核心文件

memory-workflow/
├── SKILL.md           # 本文件
├── memory_ops.py      # CLI 入口
├── README.md          # 使用说明
└── scripts/
    ├── config.py      # 配置（路径、服务地址）
    ├── store.py       # 三层存储写入
    ├── search.py      # FTS5 + Jaccard 搜索
    ├── fts5.py        # FTS5 全文索引
    ├── tools.py       # BaseTool 封装（6个工具）
    └── save_session.py # /save 命令处理器

工具接口

工具	说明	依赖
MemorySearch	语义搜索，FTS5 + 字符N-gram Jaccard 排序	无
MemoryStore	三层存储（文件 + KG + Milvus），自动去重	Ollama(KG) / Milvus(可选)
MemoryDedup	Jaccard 去重（阈值 0.85）	无
MemoryPrune	清理 N 天前记忆	无
MemoryConsolidate	合并相似记忆（Jaccard > 0.7）	无
MemoryList	列出近期记忆文件	无

搜索算法

字符级 N-gram Jaccard — 解决中文分词问题：

• "发哥在测试" → {'发哥','哥在','在测','测试'}
• 中英混合无缝支持，无需外部分词服务

搜索流程：Jaccard 召回 → FTS5 BM25 重排 → 时间衰减 rerank

数据存储路径

• 记忆文件：~/.openclaw/workspace/memory-workflow-data/memories/YYYY-MM-DD.md
• FTS5 索引：~/.openclaw/workspace/memory-workflow-data/fts5_index.db
• KG 图谱：~/.openclaw/workspace/memory-workflow-data/knowledge-graph/kg.db
• 元数据：~/.openclaw/workspace/memory-workflow-data/knowledge-graph/kg.db 内

触发规则

模式	工具
"记"、"存"、"记住"、"记一下"	MemoryStore
"搜"、"找"、"记得"、"之前说过"	MemorySearch
"去重"、"整理记忆"	MemoryDedup
"合并"、"整理一下"	MemoryConsolidate
"删除"、"清理旧"	MemoryPrune
"列出记忆文件"、"有哪些记忆"	MemoryList

CLI 用法

# 搜索
python memory_ops.py search --query "发哥 工作" --limit 3

# 存储
python memory_ops.py store --content "要记忆的内容" --tag work

# /save 存储 session
python scripts/save_session.py '{"messages": [...]}'

# 去重 / 整理
python memory_ops.py dedup
python memory_ops.py consolidate
python memory_ops.py prune --days 30

# 导出工具定义
python memory_ops.py register