mcp-memory-service MCP工具

2. Expose via Cloudflare Tunnel (or your own HTTPS setup)

cloudflared tunnel --url http://localhost:8765

Real-World: Self-Hosted Docker Stack with Cloudflare Tunnel

"The quality of life that session-independent memory adds to AI workflows is immense. File-based memory demands constant discipline. Semantic recall from a live database doesn't. Storing data on my own hardware while making it remotely accessible across platforms turned out to be a feature I didn't know I needed." — @PL-Peter, discussion #602

A production-tested self-hosted deployment using Docker containers behind a Cloudflare tunnel, with AuthMCP Gateway handling authentication:

Security best practices for this setup: - Use Cloudflare ZeroTrust with subnet-based access control (e.g., allow Anthropic subnets + your own IPs) - Add Client IP Address Filtering to all Cloudflare API tokens (Dashboard → My Profile → API Tokens → Edit → Client IP Address Filtering) to limit abuse if a token leaks - If using IPv6, include your IPv6 /64 network in the allowlist (Python prefers IPv6 by default) - For long-running browser sessions, request the offline_access scope during authorization to receive a rotating refresh_token (lifetime via MCP_OAUTH_REFRESH_TOKEN_EXPIRE_DAYS, default 30 days). Without this scope, access tokens are the only credential — extend MCP_OAUTH_ACCESS_TOKEN_EXPIRE_MINUTES up to 1440 (24h) if you need longer single-shot sessions. - Consider an auth proxy like AuthMCP or mcp-auth-proxy for robust session management

Agent Quick Start

```bash pip install mcp-memory-service MCP_ALLOW_ANONYMOUS_ACCESS=true memory server --http

✨ Quick Start Features

🧠 Persistent Memory – Context survives across sessions with semantic search 🔍 Smart Retrieval – Finds relevant context automatically using AI embeddings ⚡ 5ms Speed – Instant context injection, no latency 🔄 Multi-Client – Works across 25+ AI applications ☁️ Cloud Sync – Optional Cloudflare backend for team collaboration 🔒 Privacy-First – Local-first, you control your data 📊 Web Dashboard – Visualize and manage memories at http://localhost:8000 🧬 Knowledge Graph – Interactive D3.js visualization of memory relationships 🏠 Homelab Quality Scoring – Point scoring at any OpenAI-compatible endpoint (Ollama, LiteLLM, vLLM) 🔗 Entity Extraction – Auto-links @mentions, #tags, URLs, and file paths from memory content to a queryable entity graph 💡 Insight Cards – Consolidation detects patterns, trends, and knowledge gaps across your memory corpus and surfaces them as structured insights 🏷️ Tag Match Filtering – tag_match=AND/OR on memory_search for precise multi-tag queries

Homelab / self-hosted quality scoring (v10.45.0+): set MCP_QUALITY_AI_PROVIDER=openai-compatible to score memories with your local LLM instead of ONNX or a cloud API:

```bash MCP_QUALITY_AI_PROVIDER=openai-compatible MCP_QUALITY_AI_BASE_URL=http://localhost:11434/v1 # Ollama MCP_QUALITY_AI_MODEL=qwen2.5:7b-instruct

🎥 2-Minute Video Demo

Technical showcase: Performance, Architecture, AI/ML Intelligence & Developer Experience

3. In claude.ai: Settings → Connectors → Add Connector

3. Add connector in claude.ai Settings → Connectors with the tunnel URL

See [Remote MCP Setup Guide](docs/remote-mcp-setup.md) for production deployment with Let's Encrypt, nginx, and Docker.

</details>

<details>
<summary><strong>🔧 Advanced: Custom Backends & Team Setup</strong></summary>

For production deployments, team collaboration, or cloud sync:

Choose from: - SQLite (local, fast, single-user) - Cloudflare (cloud, multi-device sync) - Hybrid (best of both: 5ms local + background cloud sync) - Milvus (dedicated vector DB — Milvus Lite file, self-hosted, or Zilliz Cloud)

ℹ️ For long-lived services (MCP servers, web backends, notebook sessions), prefer Docker Milvus or Zilliz Cloud over Milvus Lite. See docs/milvus-backend.md for why.

</details>

---

MCP_QUALITY_AI_API_KEY=ollama # optional

Recommended models: `qwen2.5:7b-instruct` (Ollama), `mlx-community/Qwen2.5-7B-Instruct-4bit` (MLX), or any instruct model via LiteLLM proxy. On endpoint failure, scoring falls back to implicit signals automatically.

**Docker `:quality-cpu` tag** — for users who want the built-in local ONNX quality scoring (`ms-marco-MiniLM-L-6-v2` and `nvidia-quality-classifier-deberta`) without managing the one-time ONNX export themselves, and without shipping `torch`/`transformers` in their container:

The :quality-cpu image pre-exports both models at build time and ships only onnxruntime at runtime — no PyTorch dependency at deploy time. See tools/docker/README.md for details.

REST API running at http://localhost:8000

BASE_URL = "http://localhost:8000"

🛠️ CLI Server Lifecycle Commands

In addition to memory server --http (foreground mode), the CLI now includes server lifecycle commands for background HTTP management:

```bash

Persistent Shared Memory for AI Agent Pipelines

Open-source memory backend for AI agents — REST API, MCP, OAuth, CLI, dashboard. One self-hosted service, every transport. Agents store decisions, share causal knowledge graphs, and retrieve context in 5ms — without cloud lock-in or API costs.

Works with LangGraph · CrewAI · AutoGen · any HTTP client · Claude Desktop · OpenCode

---

---

Comparison with Alternatives

vs. Commercial Memory APIs

vs. MCP-Native Alternatives

MemPalace is an MCP-native alternative that went viral in April 2026 with strong LongMemEval claims. A community code review (Issue #27) subsequently showed that the headline numbers reflect the underlying vector store rather than the advertised Palace architecture, and the maintainers acknowledged most points. We keep the comparison here for transparency, but readers should interpret the scores with that context in mind.

Why the benchmark gap? Two independent factors:

1. Ingestion granularity. MemPalace stores each conversation as a single unit (session-level). LongMemEval asks "which session contains the answer?" — a question that session-level storage answers structurally. mcp-memory-service defaults to turn-level storage (one entry per message), which enables fine-grained retrieval ("what exactly did the user say about X?") but spreads a session's signal across many entries. Using memory_store_session (added in v10.35.0) brings our score to 86.0% R@5.
2. What the 96.6% actually measures. Per Issue #27, MemPalace's headline number is produced in "raw mode" — plain text stored in ChromaDB with default embeddings. The Palace architecture (Wings, Rooms, Halls) is not active in that configuration; "Halls" exist only as metadata strings with no effect on ranking. The 96.6% is therefore a ChromaDB + default-embedding baseline, not a measurement of MemPalace's structural retrieval features. A direct "apples-to-apples" architectural comparison is not possible with the published numbers.

¹ Measured in MemPalace "raw mode" (plain text in ChromaDB with default embeddings). Per Issue #27, the Palace structural features are bypassed in this configuration. ² 100% result uses optional LLM reranking (~500 API calls) on a partially tuned test set. Clean held-out score (as reported by the maintainers): 98.4% R@5.

---