# MCP Memory Service

> Open-source, self-hosted persistent memory backend for AI agents with REST API, MCP, knowledge graph, and web dashboard.

MCP Memory Service is an open-source memory backend maintained by doobidoo (copyright holder listed as Heinrich Krupp) that gives AI agents and assistants persistent, shared memory across sessions. It exposes one self-hosted service through a REST API, MCP, OAuth, a CLI, and a web dashboard. It is released under the Apache 2.0 license; the latest listed release is v11.15.0.

## What It Is

MCP Memory Service stores project context, decisions, and code patterns so new chats and agent runs start with prior knowledge instead of from zero. Agents write memories and retrieve them with semantic search. The README says embeddings run locally via ONNX, so memory stays on the user's own infrastructure.

## How Agents Use It

Agent pipelines such as LangGraph, CrewAI, and AutoGen talk to the REST API with plain HTTP. They store memories with `POST /api/memories`, search with `POST /api/search`, and retrieve by tag with `POST /api/search/by-tag`. An `X-Agent-ID` header tags memories by agent identity for scoped retrieval, a `conversation_id` bypasses deduplication for incremental conversation storage, and SSE events notify when memories are stored or deleted. MCP clients such as Claude Desktop and Claude Code connect via `memory server`, and a Remote MCP mode with OAuth 2.0 supports claude.ai and ChatGPT in Developer Mode.

## Knowledge Graph and Consolidation

Memories are linked in a knowledge graph with typed edges such as causes, fixes, and contradicts, visualized in an interactive D3.js view. Entity extraction links @mentions, #tags, URLs, and file paths. Autonomous consolidation compresses older memories and surfaces patterns, trends, and knowledge gaps as insight cards.

## Storage Backends and Setup

Install with `pip install mcp-memory-service`. Backends include SQLite, Cloudflare, a Hybrid mode with local reads and background cloud sync, and Milvus. By default the server binds to 127.0.0.1; exposing it to the network is advised only with authentication, firewall rules, or TLS. The web dashboard has eight tabs: Dashboard, Search, Browse, Documents, Manage, Analytics, Quality, and API Docs.

## Features
- Persistent memory across sessions with semantic search
- Framework-agnostic REST API
- MCP server support
- Knowledge graph with typed edges
- X-Agent-ID header for agent-scoped memories
- SSE events for memory changes
- Local ONNX embeddings
- Web dashboard with eight tabs
- Autonomous memory consolidation and insight cards
- Entity extraction
- Tag match filtering (AND/OR)
- Remote MCP with OAuth 2.0
- Multiple storage backends: SQLite, Cloudflare, Hybrid, Milvus
- Quality scoring via OpenAI-compatible endpoints

## Integrations
LangGraph, CrewAI, AutoGen, Claude Desktop, Claude Code, OpenCode, Gemini CLI, Codex CLI, Goose, Aider, Amp, VS Code, Cursor, Windsurf, Raycast, JetBrains, Zed, ChatGPT, claude.ai, Cloudflare, Milvus, Ollama, LiteLLM, vLLM, Home Assistant

## Platforms
WINDOWS, MACOS, LINUX, WEB, API, VSC_EXTENSION, JETBRAINS_PLUGIN, CLI

## Pricing
Open Source

## Version
v11.15.0

## Links
- Website: https://github.com/doobidoo/mcp-memory-service
- Documentation: https://github.com/doobidoo/mcp-memory-service/blob/main/docs/README.md
- Repository: https://github.com/doobidoo/mcp-memory-service
- EveryDev.ai: https://www.everydev.ai/tools/mcp-memory-service