MIRIX-2.0: Multi-Agent Memory System Hits High Accuracy Without Fine-Tuning

Wei-Chieh Huang · hf · 2026-11-16

Wei-Chieh Huang released MIRIX-2.0 on Hugging Face, a multi-agent memory system for LLM agents that achieves high accuracy and efficiency on QA and procedural benchmarks without any fine-tuning.

Per the method diagram, a Meta Memory Manager routes chat/image/task inputs to six specialized memory writers (episodic, semantic, procedural, core, resource, vault); a typed evidence plane keeps complete records with provenance down to source row IDs; a provenance hypergraph indexes facts and citations; and retrieval returns a bounded, citation-balanced evidence set per query. An AutoDream module runs outside the write path, merging same entities via an LLM verifier and distilling reusable skills from sessions with traceable edits. Paper: arXiv:2611.12068.

Original post →

More from coding & agent

coding & agent channel →