MIRIX-2.0 Memory System Tops LongMemEval-S by 14.3 Points
TL;DR
- MIRIX-2.0 hits 91.4% on LoCoMo and 72.6% on LongMemEval-S, leading the strongest baseline by 14.3 points.
- With AutoDream disabled, it gains 31.0 and 22.0 points over HyperMem on single-hop and multi-hop document QA.
- The system reaches 49.2% on MetaClaw-Bench against 44.0% for MetaClaw's skill library, needs no fine-tuning, and runs with closed and open models.
MIRIX-2.0, a multi-agent memory system for LLM agents, reaches 72.6% on LongMemEval-S and 91.4% on LoCoMo in a new arxiv preprint from Wei-Chieh Huang and co-authors. On LongMemEval-S, the authors write, it "leads the strongest baseline by 14.3 points."
The architecture is six "typed memory agents" that keep complete records, indexed by what the paper calls a "provenance hypergraph" with "n-ary frames that cite their rows." Retrieval, per the abstract, "runs under explicit budgets, so the evidence a question inspects does not grow with the history."
The larger gains show up on document QA when a secondary component called AutoDream is switched off. With it disabled, the paper reports "31.0 and 22.0 points over HyperMem on single-hop and multi-hop document QA." On procedural skills, MIRIX-2.0 "reaches 49.2% on MetaClaw-Bench, against 44.0% for MetaClaw's skill library," and posts the highest mean success among compared methods on ALFWorld and WebShop.
The pitch for practitioners sits in one flat line. The system "needs no fine-tuning, runs with closed and open models, and its code is public."
The abstract hedges one place. It reports MIRIX-2.0 is "the most accurate of the compared systems on four of five QA benchmarks" (not five of five), and names no baseline LLMs. The "faster than HyperMem on all five timed benchmarks" claim is also conditional: AutoDream has to be disabled for it to hold.
Originally reported by paper
Read the original article →Original headline: MIRIX-2.0 Multi-Agent Memory Leads Strongest Baseline by 14.3 Points—No Fine-Tuning Required