Mnemis: Dual-Route Retrieval on Hierarchical Graphs for Long-Term LLM Memory
New framework combines fast similarity search with deliberate reasoning for superior AI memory.
Deep Dive
Researchers led by Zihao Tang introduced Mnemis, a novel long-term memory framework for LLMs. It organizes data into hierarchical graphs and uses a dual-route retrieval system, integrating fast 'System-1' similarity search with a deliberate 'System-2' global selection mechanism. This approach scored 93.9 on the LoCoMo benchmark using GPT-4.1-mini, outperforming existing methods like RAG and Graph-RAG by enabling more comprehensive and contextually relevant information recall.
Why It Matters
Enables more reliable, context-aware AI assistants and agents by solving a core limitation in how LLMs remember past interactions.