Memori: Persistent Memory Layer Achieves 82% Accuracy Using Only 5% of Full Context Tokens — Outperforms Mem0, Zep, and LangMem on LoCoMo

arXiv·medium signal

Memori (arXiv 2603.19935, March 20) is an LLM-agnostic memory layer that converts unstructured dialogue into compact semantic triples and conversation summaries via an Advanced Augmentation pipeline. On LoCoMo benchmark: 81.95% accuracy using only 1,294 tokens per query (~5% of full context), beating Zep (79.09%), LangMem (78.05%), and Mem0 (62.47%). Single-hop reasoning hits 87.87%. Embeddings use Gemma-300 for retrieval.