The Intelligent Memory Backend for AI Agents

---

What is Smaran?

Smaran is a state-of-the-art memory infrastructure layer designed from the ground up for AI agents. It acts as the ultimate long-term memory engine for agents like Claude Desktop, Cursor, and custom LLM workflows, allowing them to instantly recall your knowledge with perfect context.

Traditional Retrieval-Augmented Generation (RAG) systems suffer from "context bloat"—repeated facts, messy data, and degraded LLM reasoning. Smaran solves this.

When you plug Smaran into your agent, it acts as a hyper-intelligent brain: intercepting raw context, filtering duplicates mathematically on the client side, and executing millisecond edge-side reranking before serving only the highest-quality, distilled facts back to the LLM.

---

Core Features

• Edge Reranking via Cloudflare Workers AI: Employs a blazing-fast Cross-Encoder (@cf/baai/bge-reranker-base) directly at the network edge, ensuring low-latency intelligence routing.
• Client-Side WASM Deduplication: Filters exact and near-exact duplicates before network transport using an ultra-fast WebAssembly port of the all-MiniLM-L6-v2 model.
• Graph-Based Consolidation: Uses the Louvain Modularity algorithm compiled to WASM to detect dense clusters in your memory space, automatically synthesizing messy notes into unified facts.
• Universal MCP Integration: Connects effortlessly to Claude Desktop via the standard Model Context Protocol (MCP).

---

Getting Started

1. Get an API Key (Closed Beta)

Smaran is currently in a strict closed beta to ensure we provide a premium, high-quality onboarding experience.

Join the waitlist at smaran.ai to request access. We review every request personally and will email you an API key if you're a good fit.

2. Test the Edge Reranker Locally (No API Key Required)

If you want to witness the power of our Cloudflare Edge Reranker right now without an API key, we have provided a mocked End-to-End test script. This script intercepts a core API call, mocks 50 messy baseline memories, and pipes them directly to the live Smaran Edge Reranker Worker.

# 1. Install dependencies across the monorepo
bun install

# 2. Run the E2E test script
cd apps/mcp
bun run test-mcp-e2e.ts

Watch your terminal as the Cloudflare Worker rescues the target memory from the absolute bottom of the pile (position 43) all the way up to position 1.

3. Connect to Claude Desktop (Requires API Key)

Once you have secured your API key, you can give Claude Desktop a perfect memory using our Model Context Protocol (MCP) server.

1. Start the local MCP server:

   cd apps/mcp
   npx wrangler dev --port 8788

2. Configure Claude Desktop: Open your Claude configuration file (claude_desktop_config.json) and register the Smaran server:

   {
     "mcpServers": {
       "smaran": {
         "command": "npx",
         "args": ["@modelcontextprotocol/inspector", "http://127.0.0.1:8788/mcp"]
       }
     }
   }

   (Note: Because OAuth proxying is currently limited in some MCP clients, manual authentication via the inspector may be required on your first run.)

---

Architecture & Repository Structure

Smaran is built as a high-performance Turbo Monorepo, separating core math engines, cloud infrastructure, and client applications.

Applications (apps/)

• web: The primary Smaran dashboard and landing page (Next.js).
• mcp: The Model Context Protocol server, deployed on Cloudflare Workers for global low latency.
• reranker-worker: The specialized Cloudflare Worker executing the BAAI Cross-Encoder model.
• browser-extension & raycast-extension: Frontend tools for capturing context directly into your Smaran graph.

Core Packages (packages/)

• client-dedup-wasm: Rust-compiled WebAssembly engine for zero-latency client-side duplication filtering.
• memory-consolidator: Rust-compiled Louvain Modularity engine for graph clustering.
• benchmark: Automated testing harness for evaluating RAG retrieval accuracy.
• tools: Built-in integrations for Slack, Notion, Google Drive, and OneDrive.

---

Manual Build Instructions

If you are contributing to the core mathematical WASM packages, you will need to compile them manually:

1. Install the Rust Toolchain (Ensure the wasm32-unknown-unknown target is installed).
2. Build the WASM packages:

   cd packages/wasm-math
   wasm-pack build --target bundler --release
   
   cd ../memory-consolidator
   wasm-pack build --target bundler --release

---

License & Copyright

Smaran is proprietary software. All rights reserved. For inquiries, contact ayush@smaran.ai.