claude-rag

Never lose context. Ever.

The first open source RAG plugin for Claude Code.
Every conversation, tool call and AI response — indexed and searchable across all your sessions.

Website  |  Install  |  Features  |  Commands  |  MCP Tools  |  Issues

---

Why?

65% of developers cite context loss as their #1 frustration with AI coding assistants. Claude Code forgets everything between sessions. Context compaction destroys your instructions mid-conversation. The same mistakes happen again and again.

Claude RAG fixes this. It captures every interaction in the background, embeds it with Gemini, and injects relevant context into every new prompt — automatically.

What makes it different

	Claude RAG	claude-mem	Cursor Memory
Embeddings	Gemini 1536d	MiniLM 384d	Proprietary
Search	Hybrid (Vector + BM25 + RRF)	ChromaDB	N/A
Vector index	DiskANN (scales to millions)	ChromaDB	N/A
Multimodal	Images, PDF, audio, video	Text only	Text only
Cost to user	Free	2x API cost (SDK agent)	Included
Architecture	Cloud (works across devices)	Local only	Local only
Turn linking	Q + Tools + A grouped	Observations	N/A
Privacy tags	<private> stripping	<private>	Content exclusion
License	MIT	AGPL-3.0	Proprietary
Post-compaction recovery	Auto re-inject context	No	No
Tech debt tracking	Built-in	No	No
Decision journal	Auto-tagged	No	No

Installation

# 1. Add the marketplace
/plugin marketplace add ThisisYoYoDev/claude-plugins

# 2. Install the plugin
/plugin install claude-rag@yoyodev

# 3. Create your free account
/claude-rag:setup

# 4. Done — start coding normally

The setup auto-detects your email from git config, sends a magic link, and configures everything. Your conversations are captured automatically from that point on.

Already have an account? Use /claude-rag:login to sign in on a new machine.

How it works

 You type a prompt
     │
     ├─ 1. RAG searches past sessions (hybrid: vector + BM25)
     │     → relevant context injected silently into your prompt
     │     → continuation context: "where you left off" from last session
     │
     ├─ 2. Claude responds using your full history as context
     │
     └─ 3. Everything is captured in the background
           → prompt, tools, response grouped by turn
           → embedded with Gemini 1536d
           → stored in PostgreSQL + DiskANN
           → searchable across all sessions forever

What you see at startup

⎿ Claude RAG v0.3.1 — resumed + 3 contexts from "myproject"

What you see on each prompt

❯ how did we fix the auth bug?

⎿ 🔍 RAG found 3 matches (8ms)
    💬 87% — Q: Fix the authentication bug in login
       🔧 2x Read, 1x Bash
       🤖 A: Changed TOKEN_EXPIRY from 3600 to 86400... (247 tokens)
    💬 72% — Q: Do the auth tests pass?
       🤖 A: 14 passed, 0 failed (5 tokens)

● The auth bug was in `src/auth/token.ts`...

Features

Capture & Memory

• Auto-capture — every prompt, AI response, tool call, and sub-agent activity
• Turn linking — Q + Tools + A grouped together via transcript promptId
• Token tracking — real Anthropic token counts (input, output, cache) per session
• Session summaries — auto-generated at session end, cached for fast retrieval
• Continuation context — "where you left off" injected at session start
• Post-compaction recovery — re-injects critical context when Claude's memory is compressed
• Multimodal — images, PDFs, audio, video uploaded to S3 and embedded with Gemini

Search & Intelligence