MIRIX-2.0 Memory System Passes Accuracy Tests

Grafy Research Lab has released MIRIX-2.0, a multi-agent memory system, and the news matters because it tackles a central weakness in AI agents: how to preserve useful long-term memory without losing precision or creating unmanageable storage. According to the research team, the system is intended to be practical, accurate and efficient at once. The development is from Grafy Research Lab, and the source record includes a feed summary and an arXiv paper. The team describes long-term memory as a way for LLM agents to personalize, stay consistent and reuse prior work. The source also notes that parametric memories need training, consolidation can lose details for good, and keeping every record can leave a growing store to search.

What Is MIRIX-2.0?

MIRIX-2.0 is described as a multi-agent memory system. The source explains that long-term memory can help large language model agents personalize behavior, stay consistent over time, and reuse what worked before. The researchers also outline problems with parametric memories, which may require training, and with consolidation processes that can cause details to be lost permanently. Keeping every record, the source adds, can lead to a growing store that is hard to search.

The team presented MIRIX-2.0 as a design for practical memory. The original source is the Grafy Research Lab paper, and the feed summary frames it around accuracy, efficiency and practical deployment concerns. The report does not claim that the system removes every memory problem. Instead, it positions MIRIX-2.0 as a multi-agent approach to long-term memory. The source material focuses on long-term memory as a way to support personalization, consistency and reuse, while also acknowledging the technical limits of parametric memory and consolidation.

How Did MIRIX-2.0 Perform?

The most concrete claim in the source concerns benchmark performance. According to the feed summary, MIRIX-2.0 was the most accurate of the compared systems on four of five question-answering benchmarks. It reached 91.4 percent on LoCoMo and 72.6 percent on LongMemEval-S, the source said it led the strongest baseline by 14.3 points. When AutoDream was disabled, the source reports that MIRIX-2.0 gained 31.0 and 22.0 points over HyperMem on the reported comparisons.

The source lists the authors as Wei-Chieh Huang, Weizhi Zhang, Jiayan Nan, Peng Xia and Yu Wang. It also notes that code is available at the source URL. The numbers are significant because memory systems often trade off between remembering more, retrieving relevant information and keeping accuracy high. MIRIX-2.0's reported results suggest strong performance on at least some long-context and long-term memory questions. The source does not say the results are final or that they apply to every agent use case.

What Is AutoDream?

AutoDream appears in the source as part of the MIRIX-2.0 discussion. The available material does not provide a full technical explanation of AutoDream, but it does connect the method to the benchmark results. The source states that with AutoDream disabled, MIRIX-2.0 gained 31.0 and 22.0 points over HyperMem on the reported comparisons.

That detail matters because it offers an ablation. It implies that AutoDream may have influenced the system's behavior or evaluation. The source does not specify how AutoDream works or why its removal changed the outcome. Readers interested in the technical mechanism would need to consult the full paper, since the provided source material is limited. The AutoDream detail is therefore presented as an important part of the reported findings, even if its function is not fully explained in the summary.

What Does This Mean for AI Agents?

For AI agents, the practical question is memory. A useful long-term memory system might help an agent remember preferences, decisions and prior successes. The source says parametric memories need training and consolidation can lose details for good. It also warns that keeping every record can create a growing store to search.

MIRIX-2.0's reported accuracy results are early but notable. The source says the system was designed to be practical, accurate and efficient. Code availability could support further testing. The practical implication is not that every agent should adopt the system immediately, but that multi-agent memory architectures may become a focal point for evaluation.

Grafy's Graph product tracks prompts, outputs and model history in a branching structure, which is a close conceptual neighbor to agent memory. A reader who wants to explore how context and changes are kept across a workflow can look at Grafy Graph. For direct experimentation with multiple models in a chat context, Grafy Chat may also be useful.

Frequently asked questions

How accurate was MIRIX-2.0 on the reported benchmarks?

According to the source, MIRIX-2.0 was the most accurate of the compared systems on four of five QA benchmarks. It reached 91.4 percent on LoCoMo and 72.6 percent on LongMemEval-S, leading the strongest baseline by 14.3 points, the source said. With AutoDream disabled, it gained 31.0 and 22.0 points over HyperMem on the reported comparisons. The source did not say these were the only results.

What does AutoDream mean in the MIRIX-2.0 results?

The source names AutoDream but does not explain its role in detail. It reports that with AutoDream disabled, MIRIX-2.0 gained 31.0 and 22.0 points over HyperMem on the reported comparisons. The question asks what AutoDream means, and the source presents it as part of an ablation. The available summary does not define the method.

What should developers take from MIRIX-2.0?

The source says MIRIX-2.0 was designed to be practical, accurate and efficient. It describes long-term memory as helpful for personalization, consistency and reuse, while noting risks in parametric memory and consolidation. The source also says code is available. Developers should treat the benchmark numbers as early evidence and not as proof of real-world readiness.

---

Source: Grafy Research Lab — MIRIX-2.0: Multi-Agent Memory System