Memory‑Plus: A Local RAG Memory Store for MCP Agents

Memory‑Plus is a lightweight, locally hosted Retrieval‑Augmented Generation (RAG) server that lets MCP agents persist, query, and visualize “memories” across sessions. In practice, it acts as a personal knowledge base that your AI assistant can read from and write to while still remaining entirely offline after the initial setup. By storing user notes, ideas, and contextual snippets in a structured format, developers can give their agents a long‑term memory that survives restarts and can be queried in natural language.

The server’s core value lies in its simplicity: a single command‑line binary powered by the UV runtime, which communicates with Claude or other MCP clients via the standard MCP protocol. Once a Google API key is supplied for Gemini embeddings, Memory‑Plus builds vector representations of each memory entry. This enables fast keyword or semantic search without requiring a separate database or cloud service. The agent can then ask for the “last 10 notes” or “all memories about project X,” and Memory‑Plus returns relevant entries along with their timestamps.

Key capabilities include:

• Record & Update – Agents can append new memories or modify existing ones, preserving a full revision history so prior states are never lost.
• Retrieve & Recent – Query by keyword, topic, or simply fetch the most recent items. The vector index ensures sub‑second responses even for thousands of entries.
• Visualization – An interactive graph clusters memories by similarity, revealing hidden relationships and making it easy for developers to audit the assistant’s knowledge.
• File Import & Deletion – Bulk ingestion of PDFs or text files, and safe removal of outdated memories.
• Memory‑for‑Memories – A higher‑level resource that teaches the agent when to recall or ignore past interactions, giving fine‑grained control over its own memory usage.
• Versioning – Every update creates a new version, allowing rollback or audit of how the assistant’s knowledge evolves over time.

Typical use cases span from personal productivity tools—where a user wants an AI to remember grocery lists, meeting notes, or project ideas—to enterprise workflows that require compliance‑ready, on‑premise knowledge bases. Developers can embed Memory‑Plus into a larger MCP ecosystem, pairing it with other tool servers (e.g., file managers or databases) to create fully autonomous agents that can reason, plan, and act based on a persistent internal state. The server’s integration is straightforward: it exposes standard MCP resources and tools, so any compliant client can invoke memory operations with simple JSON payloads.

In summary, Memory‑Plus gives developers a zero‑cost, privacy‑preserving RAG layer that turns an otherwise stateless AI assistant into a persistent, context‑aware partner. Its combination of vector search, versioning, and visual analytics makes it an indispensable component for building reliable, long‑running AI workflows.