Qdrant MCP on Raspberry Pi 5

接入你的智能体
从发布者处安装。不是 PlugMCP 端点。
远程 URL 还是本机命令? 两种安装方式的区别
Persistent semantic memory for AI agents — local, private, $0 cost.
Give your AI agent a real memory that survives reboots, searches by meaning (not keywords), and runs entirely on a Raspberry Pi 5 with no cloud dependencies.
New: Includes an OpenClaw Hard Enforcement Plugin that automatically injects memories before every response — no LLM decision required.
Note (2026-03-19): drift-memory was deprecated in Feb 2026 — all 105 entries migrated into Qdrant. This repo now stands alone. Embeddings migrated from all-MiniLM-L6-v2 (384d) → Gemini gemini-embedding-001 (3072d) via OpenClaw's memorySearch config.
AI agents forget. Between sessions, context compaction loses details, and text-based memory search only finds exact keyword matches. A typical agent can make 10+ memory errors in its first week — wrong dates, stale data, presenting old topics as new.
Local vector database (Qdrant) with MCP (Model Context Protocol) integration. Facts are stored as 384-dimensional embeddings and retrieved by semantic similarity — meaning-based, not keyword-based.
"What's the wifi password?" → finds "Office wifi: network 'Workspace5G', password is 'correct-horse-battery'"
The words don't match, but the meaning does. That's the difference.
# Qdrant MCP Server
pip3 install mcp-server-qdrant
# mcporter (MCP client)
npm install -g mcporter
mkdir -p ~/.mcporter
Create ~/.mcporter/mcporter.json:
{
"mcpServers": {
"qdrant-memory": {
"command": "mcp-server-qdrant",
"description": "Persistent vector memory using local Qdrant storage",
"env": {
"QDRANT_LOCAL_PATH": "/home/pi/.qdrant-data",
"COLLECTION_NAME": "agent-memory"
}
}
}
}
Note: Use an absolute path for
QDRANT_LOCAL_PATH. Tilde (~) is not expanded in environment variables passed by mcporter. Ifmcporter listcan't find the server, also try the full path forcommand(e.g./home/pi/.local/bin/mcp-server-qdrant).
# Check server is registered
mcporter list
# Store a memory
mcporter call qdrant-memory.qdrant-store \
information="The project runs on a Raspberry Pi 5 in my office"
# Search by meaning
mcporter call qdrant-memory.qdrant-find \
query="Where does the project run?"
That's it. No Docker, no cloud accounts, no GPU.
Agent wants to remember something
↓
mcporter call qdrant-memory.qdrant-store information="..."
↓
mcp-server-qdrant spawns (STDIO)
↓
all-MiniLM-L6-v2 creates 384-dim embedding (ONNX, CPU)
↓
Qdrant stores vector + text in local SQLite
↓
Process exits cleanly
Agent wants to recall something
↓
mcporter call qdrant-memory.qdrant-find query="..."
↓
Query embedded with same model
↓
Cosine similarity search over stored vectors
↓
Top results returned by semantic relevance
Key insight: The MCP server is stateless — it spawns per-call, does its work, and exits. No daemon process eating RAM. Qdrant's local mode uses SQLite, so data persists without a running server.
Use seed-memory.sh to bulk-load memories:
./seed-memory.sh
Or store individual facts:
mcporter call qdrant-memory.qdrant-store \
information="Weekly standup is every Monday at 09:00 in the main conference room" \
metadata='{"type": "recurring", "day": "monday"}'
The optional metadata field accepts any JSON — useful for filtering or categorization.
Store decisions, corrections, and facts. Before generating responses, search memory to verify claims. Prevents confabulation.
"Where did I put the spare keys?" / "What was the name of that restaurant?" — semantic search over your life notes.
Second collection for household or team info — schedules, appointments, decisions. Query via chat bot or voice assistant.
Store architecture decisions, meeting notes, requirements. New team members can query: "Why did we choose Postgres over MongoDB?"
{
"mcpServers": {
"work-memory": {
"command": "mcp-server-qdrant",
"env": {
"QDRANT_LOCAL_PATH": "/home/pi/.qdrant-data",
"COLLECTION_NAME": "work"
}
},
"personal-memory": {
"command": "mcp-server-qdrant",
"env": {
"QDRANT_LOCAL_PATH": "/home/pi/.qdrant-data",
"COLLECTION_NAME": "personal"
}
}
}
}
If you prefer containerized Qdrant instead of local SQLite mode:
# k3s / Kubernetes deployment
apiVersion: apps/v1
kind: Deployment
metadata:
name: qdrant
spec:
replicas: 1
selector:
matchLabels:
app: qdrant
template:
metadata:
labels:
app: qdrant
spec:
containers:
- name: qdrant
image: qdrant/qdrant:latest
ports:
- containerPort: 6333
volumeMounts:
- name: qdrant-storage
mountPath: /qdrant/storage
volumes:
- name: qdrant-storage
persistentVolumeClaim:
claimName: qdrant-pvc
---
apiVersion: v1
kind: Service
metadata:
name: qdrant
spec:
selector:
app: qdrant
ports:
- port: 6333
targetPort: 6333
Then point mcporter at the server instead of using local mode:
{
"mcpServers": {
"qdrant-memory": {
"command": "mcp-server-qdrant",
"env": {
"QDRANT_URL": "http://qdrant:6333",
"COLLECTION_NAME": "agent-memory"
}
}
}
}
Note: You cannot use both
QDRANT_URLandQDRANT_LOCAL_PATHsimultaneously. UseQDRANT_URLfor server mode (Docker/k3s),QDRANT_LOCAL_PATHfor embedded local mode (no server needed).
mcpServers in its config schema (gateway crash-loops). Use mcporter as the bridge — it works seamlessly via the bundled mcporter skill.The basic setup above requires the LLM to decide to query Qdrant. In practice, agents "forget" to use memory tools — they're optional, and LLMs skip them when rushing to answer.
Hard Enforcement solves this: an OpenClaw plugin that automatically queries Qdrant before every response and injects relevant memories into the prompt. No LLM decision required.
User message arrives
↓
before_agent_start hook fires
↓
Plugin extracts user query
↓
mcporter calls qdrant-find
↓
Top 5 results injected as prependContext
↓
LLM sees memories BEFORE generating response
cd openclaw-plugin
npm install
npm run build
# Copy to OpenClaw extensions
sudo cp -r . /usr/lib/node_modules/openclaw/extensions/nox-memory-qdrant/
# Restart OpenClaw
sudo systemctl restart openclaw-gateway
openclaw-plugin/
├── package.json # @nox/memory-qdrant
├── openclaw.plugin.json # Plugin manifest
├── tsconfig.json
└── src/
├── index.ts # Plugin entry + register()
├── qdrant-client.ts # mcporter CLI wrapper
└── auto-recall.ts # before_agent_start hook
When a user asks "Wer ist Erika Mustermann?", the plugin:
mcporter run qdrant-memory qdrant-find --query "Wer ist Erika Mustermann?"## QDRANT MEMORY RECALL (automatisch, Hard Enforcement)
Relevante Informationen aus der Vektordatenbank:
1. [87%] Erika Mustermann ist Head of Data bei Acme GmbH (Quelle: people.md)
2. [72%] Erika arbeitet eng mit dem CDO Max Mustermann zusammen (Quelle: org.md)
---
Diese Informationen wurden automatisch abgerufen.
The LLM sees this context before it starts generating — no decision to "use memory" needed.
The plugin has minimal config:
{
"enabled": true
}
Skip patterns (short messages like "hi", "ok", "danke") are handled automatically to avoid unnecessary queries.
Use both together for maximum coverage: hard enforcement catches everything, soft enforcement teaches the LLM to cite sources properly.
Qdrant excels at semantic retrieval — finding facts by meaning. But it doesn't track how the agent uses knowledge over time.
drift-memory complements Qdrant with co-occurrence tracking:
┌─────────────────────────────────────────────────────────────┐
│ HYBRID MEMORY STACK │
├─────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────────────┐ ┌─────────────────────┐ │
│ │ Qdrant │ │ drift-memory │ │
│ │ (WHAT) │ │ (HOW) │ │
│ │ │ │ │ │
│ │ • 384-dim vectors │ │ • YAML files │ │
│ │ • Cosine similarity│ │ • Co-occurrence │ │
│ │ • Hard enforcement │ │ • Biological decay │ │
│ │ • ~3s latency │ │ • <100ms latency │ │
│ └─────────────────────┘ └─────────────────────┘ │
│ ↓ ↓ │
│ "What do I know "How does the user │
│ about X?" prefer to work?" │
│ │
└─────────────────────────────────────────────────────────────┘
From the drift-memory maintainers:
"Co-occurrence tells you HOW the agent uses its knowledge. Vector embeddings tell you WHAT is semantically similar. They serve different purposes."
Memories retrieved together become linked. Over sessions, patterns emerge organically — no manual relationship tagging needed.
MIT
Persistent semantic memory for AI agents on Raspberry Pi 5 — local Qdrant + MCP, no cloud, ~3s per query
工具
暂无工具列表。远程服务器在首次健康检查后会获得工具列表。
常见问题
在本页复制适用于 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 配置。它使用发布者的安装方式,而不是 PlugMCP 代理。
此记录未列出必需的环境变量。
Qdrant MCP Pi5 通过 stdio 在本地运行。
上次检查时间:2026-10-28。 Persistent semantic memory for AI agents on Raspberry Pi 5 — local Qdrant + MCP, no cloud, ~3s per query