Best MCP Servers for RAG in 2026

2026-08-18 · by bixzell

Best MCP Servers for RAG in 2026

RAG (retrieval-augmented generation) is where the MCP ecosystem is growing fastest. It’s also the category where “best” is genuinely situational: the right pick depends on whether you’re searching code, recalling past conversations, or retrieving from a document vault.

This guide splits the field into four groups — code intelligence, agent memory, knowledge bases, and video search — so you can pick by job, not by stars.

How to pick a RAG MCP server

Before the list, three filters that matter more than star count:

  1. What are you retrieving from? Code, chat history, PDFs, and video all need different indexing. A memory server won’t search your repo, and a code search server won’t remember your meetings.
  2. Local vs hosted. Privacy-sensitive teams should bias toward local-first servers (SQLite / self-hosted), which keep embeddings on your machine.
  3. Client support. Make sure the server runs in your client — Claude Desktop, Cursor, and Codex all speak MCP, but transport options differ.

1. Code intelligence & semantic code search

DeusData/codebase-memory-mcp

The category leader by a wide margin: a high-performance code intelligence MCP server that indexes a codebase into a persistent knowledge graph. It supports 158 languages, answers symbol-level queries in sub-ms, and claims 99% fewer tokens than stuffing files into context. If your agent spends more time reading code than writing it, start here.

MinishLab/semble

Semble does semantic code search for agents using embeddings and retrieval, positioning itself against grep+read with dramatically fewer tokens. It’s a focused alternative if you want search over a full knowledge graph.

  • Stars: 5,893⭐
  • Best for: agent-friendly search without a heavyweight index

bgauryy/octocode

A semantic search platform built for AI agents — find, understand, and reuse code across a project or workspace. A good middle option between a full knowledge graph and plain grep.

  • Stars: 913⭐
  • Best for: code research across many repos

2. Persistent agent memory

Memory servers are RAG over your conversation and work history: they store, deduplicate, and recall context across sessions. This is the fastest-growing slice of the category.

agentic-box/memora

Memora gives agents persistent, collective memory with deduplicating absorb, supersession lineage, semantic search, and a graph UI. It speaks MCP out of the box and is built for teams that want shared, queryable agent memory.

  • Stars: 664⭐
  • Best for: shared agent memory with a visual graph

CodeAbra/iai-personal-memory-engine

A “cyber brain for your AI” — free, local, MIT-licensed, and compatible with Cursor, Claude Code, Codex, OpenClaw and more. It tracks what you actually work on and recalls it when context runs out.

  • Stars: 658⭐
  • Best for: personal memory across many AI tools

TeleAI-UAGI/telemem

A high-performance drop-in replacement for Mem0 with semantic deduplication, long-term dialogue memory, and multimodal video reasoning. If you already have a Mem0 workflow and want something faster, this is a strong swap.

  • Stars: 484⭐
  • Best for: replacing Mem0 without re-architecting

buildingjoshbetter/TrueMemory

TrueMemory does automatic capture and recall, 100% local, in a single SQLite file with zero cloud. Works with Claude Code, Cursor, Codex, and Gemini CLI.

  • Stars: 373⭐
  • Best for: zero-cloud, one-file memory

Lyellr88/marm-memory

A local-first 3-in-1 memory layer that fuses session history, codebase indexing, and concept graphs in SQLite — supporting multi-agent swarms. The most feature-dense local option here.

  • Stars: 334⭐
  • Best for: one memory layer across code, chat, and concepts

3. Knowledge base & document retrieval

nduckmink/arkon

A self-hosted enterprise AI knowledge hub and MCP server — ingest documents, index them, and let your agents query the knowledge base. For teams that need document RAG on their own infrastructure.

  • Stars: 1,226⭐
  • Best for: self-hosted team knowledge bases

SylphxAI/pdf-reader-mcp

Gives your agent eyes for PDFs — extracting structured text, tables, and layout so the content becomes RAG-ready. The right entry point when your knowledge base is a pile of PDFs.

  • Stars: 894⭐
  • Best for: PDF-heavy document libraries

aaronsb/obsidian-mcp-plugin

A high-performance MCP server for Obsidian that gives AI tools direct vault access through semantic operations and HTTP transport. If your knowledge already lives in Obsidian, this makes it queryable without moving anything.

  • Stars: 455⭐
  • Best for: RAG over your existing Obsidian vault

4. Video & multimodal search

NVIDIA-AI-Blueprints/video-search-and-summarization

NVIDIA’s reference architecture for video analytics agents: GPU-accelerated, uses VLMs and RAG for real-time verified alerts, visual Q&A, and automated reporting. Overkill for chat history, essential for video-heavy workloads.

  • Stars: 1,805⭐
  • Best for: long-video understanding at scale

The decision tree

  • Your agent can’t find code → codebase-memory-mcp or semble
  • Your agent forgets conversations → memora or iai-personal-memory-engine
  • You need to migrate off Mem0 → telemem
  • Your knowledge is PDFs or an Obsidian vault → pdf-reader-mcp or obsidian-mcp-plugin
  • You want everything in one local file, zero cloud → TrueMemory or marm-memory
  • You process hours of video → NVIDIA video-search blueprint

Two of these servers deserve a closer look before you commit — FastMCP vs mcp-use matters if you’re building your own retrieval pipeline, and our FastMCP review covers the framework most RAG servers are built on.