Local Vector Memory

Other

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

Install

openclaw skills install local-vector-memory

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