Mnemis: Dual-Route Retrieval on Hierarchical Graphs for Long-Term LLM Memory

Mnemis is a sophisticated implementation of the 'System 1/System 2' cognitive framework applied to LLM memory. By combining standard similarity-based retrieval (System 1) with a structured, hierarchical graph traversal (System 2), it addresses the 'lost in the middle' and global reasoning failures of traditional RAG. Quantitatively, the 12 forks within just 7 days despite 0 stars strongly suggest a 'paper-with-code' release that is being immediately audited by other researchers or labs, indicating high topical relevance. However, the project faces extreme frontier risk: OpenAI (Mem0/native memory), Microsoft (GraphRAG), and Google (Gemini long-context) are all aggressively solving the same memory bottlenecks. The moat is purely algorithmic; while the dual-route approach is clever, it lacks the data gravity or proprietary infrastructure to prevent it from being absorbed as a standard design pattern in larger frameworks like LangChain or LlamaIndex. Its defensibility is currently limited to its status as a reference implementation for a specific paper.