surreal-mem-mcp

A standalone Edge-RAG Memory Daemon utilizing HTTP/SSE MCP transport. The architecture consists of a high-performance Rust core (SurrealDB/RocksDB + Axum server) and a zero-config Python CLI Bootstrapper. It bypasses heavy LLM-as-a-Judge rerankers in favor of an ultra-fast sub-millisecond Bayesian Math Outer Query combining Cosine Similarity and BM25 Posterior scores.

This Model Context Protocol (MCP) server allows your AI agents (Claude Desktop, Cursor, Gemini CLI, OpenCode, Code Puppy, etc.) to use a shared global memory system, complete with standardized behavioral rules (SOUL.md and MEMORY.md), independent of your current project directory.

• Blazing Fast: Uses Rust, Axum, and native Server-Sent Events (SSE) alongside embedded SurrealDB backed by RocksDB.
• BYOM (Bring Your Own Model): Compatible with local OpenAI-format embeddings (Ollama, LM Studio) natively.
• Bayesian Edge-RAG: Blends vector::similarity::cosine() * 0.7 and BM25 * 0.3, weighted by an epistemic prior (time decay, graph density, access counts). Let math do the parsing, not latency-heavy LLMs.
• Global Behavioral Rules: Generates ~/.surreal-mem-mcp/rules/ accessible universally via resources/list. Memory rules dynamically persist across completely disparate agent ecosystems.
• Enterprise Semantic Redaction: Intercepts and scrubs API keys (OpenAI, AWS, GCP, Anthropic, Stripe) and PII natively in Rust before they are written to the database, ensuring your local RocksDB volume is 100% leak-proof.
• Automated Bootstrapper: Ships with a smart Python script that probes inference endpoints and injects system configs dynamically.

> 💡 Curious about the technical tradeoffs? Read [ARCHITECTURE.md](./ARCHITECTURE.md) to see how our Epistemic Math Queries drastically drop LLM token consumption, and why our embedded Graph Database outperforms traditional SQLite for dense code reasoning.

We recommend u