MCP Memory Graph

What mcp-memory-graph does

The mcp-memory-graph MCP server provides persistent memory for Claude Code and other MCP clients. Memories are stored in a SQLite database on the user's machine, so decisions, notes, and other knowledge remain available after a conversation ends. The project is intended for developers and small teams that want local storage instead of a hosted memory service.

It supports semantic search as well as keyword search. A query can find related information even when the stored text uses different wording. Memories can also be connected to entities and to one another through a knowledge graph, which supports questions that require following relationships between stored items.

The server includes a local web dashboard for browsing and searching the memory store. The dashboard exposes memory counts, scopes, types, tags, quality scores, read frequency, and recent entries.

How it works

When a memory is saved, a local MiniLM-based embedding model converts its text into a 384-number vector. The server stores that vector and a keyword-search index in the same SQLite database. By default, the database is located at ~/.mcp-memory/memory.db.

Search combines semantic and keyword results, merges their rankings, and returns matches with confidence information. For MCP clients, a reranker is enabled by default to improve the ordering of top results; the README describes an approximate 200 ms cost for this step. The embedding model is downloaded the first time a memory tool runs and cached under ~/.cache/huggingface/.

A scheduled dream cycle handles maintenance tasks such as deduplication, rescoring, pruning, and reporting knowledge gaps. Claude Code users can also enable a Stop hook that sends the session transcript to their locally installed claude -p command for learning extraction. This hook can be disabled with review_on_stop: false.

Setup and configuration

Node.js 20 or newer is required. The package can be installed globally with npm, or run directly with npx -y mcp-memory-graph. Claude Code registration can be performed with claude mcp add, while npx mcp-memory-graph init can register the server, install capture and recall hooks, write configuration, install the usage skill, and schedule nightly cleanup. The --yes option accepts defaults, and --no-register skips automatic MCP registration.

The mcp-memory-graph MCP server can be removed with npx mcp-memory-graph uninstall. Running npx mcp-memory-graph consolidate starts consolidation manually.

Tools and capabilities

The README reports 51 available tools: 45 memory_* tools, three vault_* tools, and three core_memory_* tools. The documented capabilities include:

• Storing and retrieving persistent memories
• Hybrid semantic and keyword search
• Knowledge-graph relationships between memories and entities
• Browsing and searching through a local web dashboard
• Automatic session capture and recall through Claude Code hooks
• Quality scoring, usage tracking, deduplication, pruning, and consolidation

Limitations and notes

The server uses a single-process SQLite design. The README states that it generally tops out in the low hundreds of thousands of vectors, so larger deployments may need a different architecture. Local operation avoids cloud infrastructure, telemetry, API keys, and metered token costs, but commercial use is not covered by the free noncommercial license. Personal projects, hobbies, study, research, charity, education, and government use are listed as permitted; commercial use requires a paid license.

The optional Stop hook invokes a locally installed Claude Code command, which is an exception to the otherwise local processing path. Users who do not want transcript-based learning extraction can disable that hook.

Getting started with this mcp-memory-graph MCP server

Always refer to the official documentation for the most accurate and up-to-date information.