News
MIRIX-2.0, a multi-agent memory system, sets new standards in QA and procedural benchmarks with high accuracy and efficiency, as detailed in a new AI paper.
A new AI research paper, MIRIX-2.0: Multi-Agent Memory System, published on November 17, 2026, introduces an advanced memory system for large language model (LLM) agents. Developed by Wei-Chieh Huang, Weizhi Zhang, Jiayan Nan, Peng Xia, and Yu Wang, the system demonstrates significant improvements in accuracy and efficiency across multiple benchmarks without requiring fine-tuning. The open-source implementation is available on GitHub.
Breakthrough in Long-Term Memory for LLM Agents
MIRIX-2.0 addresses critical challenges in long-term memory for LLM agents, including the need for personalization, consistency, and efficient reuse of information. Traditional parametric memory systems often require training and risk losing details during consolidation, while maintaining every record can lead to inefficiencies in search processes. MIRIX-2.0 overcomes these limitations with a multi-agent memory architecture that integrates six specialized memory components: Core, Episodic, Semantic, Procedural, Resource, and Knowledge Vault.
Benchmark-Leading Performance
The system achieved leading results on four out of five QA benchmarks. Key highlights include:
- 91.4% accuracy on the LoCoMo benchmark, surpassing the strongest baseline by 14.3 percentage points.
- 72.6% accuracy on LongMemEval-S, also leading by 14.3 points.
- Significant gains over HyperMem on document QA tasks, with improvements of 31.0 points for single-hop and 22.0 points for multi-hop QA when AutoDream was disabled.
In procedural memory tasks, MIRIX-2.0 achieved the highest mean success rates among compared systems. It recorded a 49.2% success rate on MetaClaw-Bench, outperforming MetaClaw’s skill library, which scored 44.0%.
Technical Innovations
The system employs a provenance hypergraph to index episodic and semantic records using n-ary frames, ensuring efficient retrieval under explicit resource budgets. This design prevents the evidence base from growing uncontrollably with history. Additionally, the AutoDream feature consolidates duplicate anchors outside the write path and distills task sessions into reusable skills, enhancing both speed and efficiency.
MIRIX-2.0 operates seamlessly with both closed and open models, requires no fine-tuning, and supports multi-modal inputs such as text, images, voice, and screen captures. Its PostgreSQL-native BM25 full-text search and optional Neo4j-based graph memory further enhance its advanced search capabilities.
Open-Source Availability
The MIRIX-2.0 system is available as an open-source project on GitHub, allowing developers to explore its features and integrate them into their applications. The privacy-first design ensures that all long-term data remains stored locally, with user-controlled privacy settings.
This breakthrough in memory systems for LLM agents has the potential to redefine how AI models handle long-term memory, personalization, and procedural tasks, paving the way for more intelligent and efficient AI applications.
