Omni-SimpleMem: 23-Stage Autonomous Research Pipeline Discovers Multimodal Agent Memory Architecture

arXiv·medium signal

Researchers used AutoResearchClaw, a 23-stage autonomous pipeline, to discover Omni-SimpleMem — a unified multimodal memory framework for lifelong AI agents supporting text, image, audio, and video. The pipeline ran ~50 experiments autonomously, finding that bug fixes (+175%), architectural changes (+44%), and prompt engineering (+188%) each individually exceeded all hyperparameter tuning contributions combined. The resulting system achieves state-of-the-art on LoCoMo (F1=0.613, +47%) and Mem-Gallery (F1=0.810, +51%) using selective ingestion, hybrid FAISS+BM25 retrieval, and knowledge graph augmentation.