Found first: a primary source the press has not covered yet.
Researchers at the University of Illinois Chicago, Tongji University, the University of North Carolina, and UC San Diego have released MIRIX-2.0, a multi-agent memory system that scores 72.6% on LongMemEval-S, 14.3 points ahead of the strongest baseline, without fine-tuning any underlying model. The same system scores 91.4% on LoCoMo and posts the highest mean success on ALFWorld and WebShop among all compared memory methods.
What the source says
MIRIX-2.0 organizes memory across six typed stores: Core, Episodic, Semantic, Procedural, Resource, and Knowledge Vault. Episodic and semantic records are indexed by a provenance hypergraph in which every fact cites the source records it was drawn from, using n-ary role-typed frames rather than binary triples. A second component, AutoDream, distills completed task sessions into reusable procedural skills without overwriting existing records. On single-hop and multi-hop document QA, the system answers 83.0% and 71.0% of questions, 31.0 and 22.0 points above the strongest baseline in the M1 configuration (AutoDream disabled). On MetaClaw-Bench it reaches 49.2%, against 44.0% for MetaClaw's own skill library. ALFWorld mean success reaches 76.4% with GPT-6 Luna. The system works with closed and open models; code is public at GitHub.
Why it matters
Most memory architectures are designed for one regime: either tracking user history for personalization, or accumulating agent skills across task runs. MIRIX-2.0 handles both from a single store, which matters because production agents increasingly need to do both. The 14.3-point gap on LongMemEval-S is large enough to be meaningful in practice rather than a rounding artifact. The provenance hypergraph design also addresses a standing weakness in RAG-style memory: retrieved facts can be traced back to their source rows, which supports auditability. The document QA gains are reported with AutoDream off, so the full system's ceiling on those tasks is not yet characterized.