MIRIX-2.0 Beats Best LLM Memory Baseline by 14.3 Points, No Fine-Tuning

Found first: a primary source the press has not covered yet.

A new multi-agent memory architecture, MIRIX-2.0, scores 72.6% on LongMemEval-S, 14.3 percentage points above the strongest baseline, with no fine-tuning required. The paper, by Wei-Chieh Huang, Weizhi Zhang, Jiayan Nan, Peng Xia, and Yu Wang (affiliations not listed in the preprint), also reports 91.4% on LoCoMo and gains of 31.0 and 22.0 points over HyperMem on single-hop and multi-hop document QA tasks.

What the source says

MIRIX-2.0 runs six typed memory agents whose records are indexed by a provenance hypergraph tracking n-ary relationships between episodic and semantic entries. A component called AutoDream merges duplicate memory anchors without rewriting existing records, and distills completed task sessions into reusable skills. Retrieval operates under explicit token budgets so context does not grow unboundedly with history. On MetaClaw-Bench the system scores 49.2% against 44.0% for the competing MetaClaw skill library approach. It achieves the highest mean success rate on ALFWorld and WebShop among the memory methods compared.

Why it matters

Prior memory systems for LLM agents typically serve either user personalization or agent skill accumulation, rarely both from a single store. MIRIX-2.0 handles both without modifying the underlying model, so the same architecture works against closed API models and open-weight ones. The 14.3-point gap on LongMemEval-S is a substantial margin at a benchmark designed around realistic long-horizon recall. Code is public.