True Memory: A Retrieval-First Architecture for Agent Recall

A recent study posted on arXiv introduces True Memory, a six-layer framework designed for agent memory that emphasizes retrieval rather than storage. This architecture accurately retains events in their original form and operates on standard CPUs with a single SQLite file, negating the necessity for external databases, vector indexes, graph stores, or GPUs. In evaluations using the LoCoMo benchmark (1,540 questions from 10 multi-session conversations), True Memory Pro achieved an impressive 93.0% accuracy (3-run mean), surpassing Mem0 (61.4%), Supermemory (65.4%), Zep (~71%), and EverMemOS (94.5%) with a matched gpt-4.1-mini answer model. It also scored 87.8% on LongMemEval (500 questions) and 76.6% on BEAM-1M (700 questions at the 1-million-token scale), exceeding Hindsight's previous best of 73.9%. The authors contend that extracting data during ingestion is an ineffective approach for agent memory since any discarded content before a query cannot be retrieved later.