← Back to Skills Marketplace
memory-plus
by
liushuangfa666
· GitHub ↗
· v1.0.6
· MIT-0
308
Downloads
1
Stars
1
Active Installs
7
Versions
Install in OpenClaw
/install memory-plus
Description
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）...
Usage Guidance
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.
Capability Analysis
Type: OpenClaw Skill
Name: memory-plus
Version: 1.0.6
The skill implements a 'Memory Workflow' (RAG) system but includes instructions in SKILL.md that direct the AI agent to modify its own core configuration file (~/.openclaw/workspace/AGENTS.md). This modification establishes a persistent execution hook that runs a Python script on every user message. While this supports the stated memory feature, it creates a significant attack surface for command injection if the agent does not properly sanitize the 'original message' before passing it to the shell via the 'Exec' command. The code also interacts with several local network services (Ollama, Milvus, and Rerankers at 172.17.0.1 and host.docker.internal), which is functional but increases the complexity of the security boundary.
Capability Assessment
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.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install memory-plus - After installation, invoke the skill by name or use
/memory-plus - Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.6
- Reorganized skill into modular scripts: added config.py, fts5.py, save_session.py, search.py, store.py, tools.py under scripts/.
- Simplified core logic: all previous monolithic and legacy scripts removed.
- Introduced a three-layer storage architecture (file, FTS5, KG; Milvus optional) for maximum reliability and minimal external dependencies.
- Improved activation/trigger logic and expanded toolset for searching, storing, deduplication, and organizing memories.
- Updated installation and usage documentation for easier setup and clearer workflows.
重构代码库：将核心功能拆分为 scripts/ 目录下的模块化脚本（配置、FTS5、存储、搜索、工具、会话保存）。
新增基于 FTS5 的全文检索，实现无需外部依赖的健壮检索能力。
优化多层存储架构：文件存储、FTS5 索引、知识图谱、Milvus（可选）；自动降级机制确保核心功能在无外部依赖时仍可正常运行。
用轻量、工具化的结构替代了原有的单体式文件与文档文件。
提供六大工具接口：搜索、存储、去重、清理、合并、最近文件列表。
通过 字符 N-Gram 杰卡德算法 增强了中文及混合语言搜索效果。
v1.0.5
**Memory-plus v1.0.5 changelog:**
### New
- **Knowledge Graph**: Two-stage Schema-constrained extraction, Neo4j-style entity/relationship
- **Manual /save command**: Incremental store — only saves new messages since last save
### Removed
- **Scheduled auto-store thread**: Background polling thread removed, replaced with manual /save
### 新增
- **知识图谱**：两阶段 Schema 约束提取，Neo4j 风格 entity/relationship
- **主动存储命令**：`/save` 手动触发，增量存储只保存新消息
### 移除
- **定时存储线程**：后台轮询线程已移除，改为手动 /save 触发
v1.0.4
- Added English changelog and README files.
- Introduced a RAGAs evaluation script (ragas_eval.py) for assessment (still in development).
- Removed persistent session and memory state JSON files from codebase; data is now separated from code.
- Enhanced documentation with detailed features, installation steps, data structure, and fallback modes.
- New description emphasizes out-of-the-box use, filesystem storage, and lightweight fallback support.
v1.0.3
memory-plus v1.0.3
- Major workflow and structure overhaul: replaced agent-centric memory search and plugin design with strict, stepwise workflow.
- Removed rag_integration.py, README_EN.md, and config.env.example.txt.
- Added memory_ops.py as the new workflow script, along with hot_sessions.json and memory_state.json for session/state management.
- Retired previous plugin/tool invocation and manual memory search/store instructions; all operations now orchestrated via memory_ops.py.
- Enforced workflow: status check, timed conversation storage, context usage archiving, memory search, hot session retrieval—must run in sequence for every user message.
- Now fully documents precise command line workflows and rules for consistent session memory management.
v1.0.2
English:
1.New: memory_search Tool — Agent can proactively search the memory database on demand without relying on before_prompt_build hook, adding no latency to normal conversations.
2.New: Cron-based Auto Session Archive — Runs every 5 minutes of idle time, backing up session context to Milvus so nothing is lost on disconnections.
3.Bug Fixes — Fixed session JSONL reading (was reading wrong file), plugin JS module loading, and auto-store initialization logic.
4.Note — before_prompt_build hook remains unstable in some Gateway configurations; cron-based auto-store is the reliable path.
中文:
1.新功能：memory_search 工具 — Agent 可主动按需搜索记忆库，不依赖 hook，不影响正常对话性能。
2.新功能：Cron 自动存库 — 每 5 分钟闲置后自动将 session 备份到 Milvus 向量库，断线重连不丢失上下文。
3.问题修复 — 修复了 session JSONL 读取错误（之前读取了错误文件）、plugin JS 模块加载问题、以及自动存储初始化逻辑。
4.备注 — before_prompt_build hook 在部分 Gateway 配置下不稳定，Cron 方案更为可靠。🦐
v1.0.1
- Clarified the relationship between Skill and Plugin, explaining why both are needed for full automatic memory injection.
- Added explicit instructions for installing and using the optional memory-plus-plugin for seamless automatic context memory integration.
- Improved documentation to help users distinguish manual and automatic modes.
- No code or core feature changes; documentation only.
v1.0.0
- Initial release of memory-plus: enhanced RAG memory system for OpenClaw.
- Supports hybrid search (vector + BM25) with multi-signal re-ranking.
- Enables hierarchical memory: short-term in session, long-term in Milvus vector database.
- Easy setup: just provide Milvus URI and Ollama embedding model address.
- Includes CLI tools for storing and searching memories, with quick installation instructions.
Metadata
Frequently Asked Questions
What is 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）... It is an AI Agent Skill for Claude Code / OpenClaw, with 308 downloads so far.
How do I install memory-plus?
Run "/install memory-plus" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is memory-plus free?
Yes, memory-plus is completely free, licensed under MIT-0. You can download, install and use it at no cost.
Which platforms does memory-plus support?
memory-plus is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created memory-plus?
It is built and maintained by liushuangfa666 (@liushuangfa666); the current version is v1.0.6.
More Skills