Paper review · November 17, 2026
MIRIX-2.0: A Multi-Agent Memory System
MIRIX-2.0: Multi-Agent Memory System
Combining six types of memory agents that preserve original records, an n-ary hypergraph index that cites sources, and budget-based retrieval, the system improves both accuracy and speed on long-term memory QA and procedural memory tasks.
Problem
Long-term memory for LLM agents struggles to be practical, accurate and efficient at the same time. Extracting and compressing facts permanently loses detail. Keeping every record makes the search space keep growing. Binary triple graphs fragment facts that involve several participants. Parametric memory requires training and cannot be used with closed models.
Method
A Meta Memory Manager routes input to six types of memory-writing agents (Core, Episodic, Semantic, Procedural, Resource, Knowledge Vault). Complete records are first committed to a vector DB, which serves as the system of record. Episodic and Semantic records are indexed in an n-ary frame hypergraph that cites the source rows. Dates, prices and quantities are attached as literal values on frames. Similar events without dates are distinguished by an occurrence key. At query time, a deterministic planner sets a budget, traverses the graph from anchors that match the question, and retrieves only the cited records. AutoDream runs outside the write path. It merges duplicate entity anchors without modifying records and distills work sessions into skills.
Results
- The system scored 91.4% on LoCoMo and 72.6% on LongMemEval-S. On LongMemEval-S it was 14.3 points above the strongest baseline.
- On MemoryAgentBench document QA with AutoDream turned off, it scored 83.0% single-hop and 71.0% multi-hop, 31.0 and 22.0 points above HyperMem respectively.
- In the LoCoMo ablation, the hypergraph raised accuracy from 85.7% to 88.4% compared with flat memory, and AutoDream raised it further to 91.4% (+3.0 points).
- Used as procedural memory, it reached 76.4% on ALFWorld with GPT-6 Luna, above AWM (73.6%) and no memory (69.4%). On MetaClaw-Bench it scored 49.2%, ahead of the MetaClaw skill library (44.0%).
- With AutoDream turned off, it responded faster than HyperMem on all five benchmarks measured.
Limitations
As the authors note, for long documents, graph construction makes ingestion about 4 times slower than flat memory and also slower than the baselines. The time for the full system with AutoDream turned on was not measured. The retrieval cap refers to the amount of evidence a question checks; it does not measure latency as records grow longer. The LoCoMo numbers come from the authors' own runs.
Why it matters
The design principles of preserving originals, using the graph only as an index, and keeping stored data intact even on failure (fail-soft) directly help with trusting and tracing memory in production. Handling personalized user memory and agent skill memory in one store is also practical. The code is publicly available.
Key terms
- n-ary hypergraph
- A graph that groups the multiple entities taking part in one fact (for example, a flight) into a single frame along with their roles. Facts do not get fragmented as they do with binary triples.
- System of Record
- The authoritative store that holds the original data. Here the vector DB plays this role, and the graph is only an index pointing to it.
- AutoDream
- A memory cleanup process that runs outside the write path. It merges duplicate entities and distills reusable skills from work sessions.
- Procedural memory
- Memory in which an agent stores procedures and skills gained from past task experience and reuses them in later tasks.
Original
Authors: Wei-Chieh Huang, Weizhi Zhang, Jiayan Nan, Peng Xia, Yu Wang · Posted 2026-11-09 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.