← 返回 Skills 市场

Local Vector Memory

Name: Local Vector Memory
Author: jancong

作者 Cong Pendy · GitHub ↗ · v1.0.0 · MIT-0

cross-platform ⚠ suspicious

总下载

当前安装

版本数

在 OpenClaw 中安装

/install local-vector-memory

功能描述

Store, search, and manage local vector memories using Ollama embeddings with Qdrant, supporting Chinese and English text without cloud dependencies.

使用说明 (SKILL.md)

Local Vector Memory Skill

Zero-cloud vector memory using Ollama embeddings + Qdrant local storage.

Prerequisites

# Ollama with embedding model
ollama pull qwen3-embedding:4b

# Install the package
pip install local-vector-memory

Quick Reference

lvm init                    # Initialize database
lvm add "text to remember"  # Store a memory
lvm search "query"          # Semantic search
lvm search "query" --limit 3 --json  # Structured output
lvm stats                   # Show stats
lvm reindex --dir ~/notes   # Reindex markdown files
lvm delete "source_name"    # Delete by source

Python Library Usage

from local_vector_memory.core import LocalVectorMemory

lvm = LocalVectorMemory()  # uses env defaults
lvm.add("OpenClaw baseUrl must not end with /v1")
results = lvm.search("how to configure ollama")
for r in results:
    print(f"[{r['score']}] {r['source']}: {r['text'][:100]}")

Configuration

Env Var	Default	Description
`LVM_OLLAMA_URL`	`http://localhost:11434`	Must be localhost (SSRF protected)
`LVM_MODEL`	`qwen3-embedding:4b`	Embedding model
`LVM_DIMS`	`2560`	Vector dimensions
`LVM_DB_PATH`	`~/.local-vector-memory/qdrant`	Storage path
`LVM_CHUNK_SIZE`	`400`	Chunk size in chars
`LVM_CHUNK_OVERLAP`	`50`	Overlap between chunks

Embedding Model Selection

Model	Dims	Size	Chinese Hit Rate	Best For
`qwen3-embedding:4b`	2560	~2.5GB	100%	Chinese/English mixed
`bge-m3`	1024	~570MB	40%	Multilingual, low RAM
`nomic-embed-text`	768	274MB	30%	English-only, minimal RAM

Integration Patterns

With OpenClaw

Add to HEARTBEAT.md or cron for periodic reindexing:

lvm reindex --dir ~/.openclaw/workspace/memory

As a backup search layer

When memory_search doesn't find what you need:

lvm search "query" --json

Security

Ollama URL restricted to localhost only (SSRF protection)
Path traversal blocked in reindex glob patterns
Input length limits enforced (100K text, 10K query)
All data stored locally, no network calls except to local Ollama

Links

安全使用建议

This skill appears to do what it says (local vector memory using Ollama + Qdrant) but it relies on installing a third-party PyPI package and reindexing local directories, which can expose sensitive files if misused. Before installing or running: 1) verify the PyPI project and GitHub repo (maintainer, recent activity, inspect source) rather than blindly pip installing; 2) run installation inside an isolated environment (virtualenv/container) and review what files the package writes; 3) ensure LVM_OLLAMA_URL is bound to localhost and not reachable remotely; do not set it to a public endpoint; 4) restrict reindex targets (explicit directories you control) and avoid system, home dotfiles, or secret stores; 5) prefer running 'lvm' manually first to confirm behavior rather than enabling automated reindex/cron; and 6) if you need stronger assurance, review the package code or run it in a sandboxed VM before allowing the agent to invoke it autonomously.

能力评估

✓ Purpose & Capability

Name/description match the SKILL.md: it documents using Ollama embeddings and a Qdrant datastore, provides CLI usage (lvm) and a Python API, and lists relevant config env vars. Minor inconsistency: registry metadata shows no homepage/source but SKILL.md references PyPI and GitHub; this is plausible but worth verifying.

⚠ Instruction Scope

SKILL.md instructs indexing local directories (e.g., 'lvm reindex --dir ~/notes' and recommending cron/HEARTBEAT.md reindex of ~/.openclaw/workspace/memory). Reindexing arbitrary paths is intrinsic to the skill but can read and store sensitive local files if misconfigured. The doc also claims SSRF protection by restricting Ollama URL to localhost, but that is a declarative claim — the skill provides no enforcement mechanism (it's an instruction-only skill).

ℹ Install Mechanism

No install spec in the registry (instruction-only). The SKILL.md instructs 'pip install local-vector-memory' and 'ollama pull qwen3-embedding:4b'. Installing a third-party PyPI package is expected for this functionality but introduces supply-chain risk: the package and its code should be reviewed/trusted before pip install. Downloading large embedding models via 'ollama pull' is normal for local embedding but depends on the Ollama provider/source.

ℹ Credentials

The skill requires no credentials in the registry. SKILL.md lists several configuration env vars (LVM_OLLAMA_URL, LVM_DB_PATH, etc.) which are reasonable for a local vector store. However, if LVM_OLLAMA_URL is changed from localhost to an external endpoint, embeddings/contents could be transmitted off-host — the documentation's 'localhost only' guidance must be enforced by the operator.

✓ Persistence & Privilege

always:false (no forced presence) and normal model invocation are set. The skill suggests adding periodic reindexing (cron/heartbeat), which is expected for a memory tool but increases exposure if misconfigured; this is a usage risk rather than a privilege mis-declaration in the skill metadata.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install local-vector-memory
安装完成后，直接呼叫该 Skill 的名称或使用 /local-vector-memory 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v1.0.0

Initial release - Ollama + Qdrant local vector memory

元数据

Slug local-vector-memory

版本 1.0.0

许可证 MIT-0

累计安装 1

当前安装数 1

历史版本数 1

常见问题