MIRIX-2.0 keeps records verbatim, indexes with a provenance hypergraph, tops 4 of 5 QA benchmarks

MIRIX-2.0: Multi-Agent Memory System

Wei-Chieh Huang, Weizhi Zhang, Jiayan Nan, Peng Xia, Yu Wang

cs.CL

2026-11-09

Six typed memory agents plus a provenance-hypergraph index reach 91.4% on LoCoMo and lead the strongest baseline by 14.3 points on LongMemEval-S, with no fine-tuning.

What problem this solves

LLM agents that work across sessions need memory that is accurate, practical, and efficient at once. Accurate means a name, a date, a count, or the line between two similar events survives storage. Practical means no training, works with closed models, traceable answers, and consolidation that cannot destroy what was stored. Efficient means answer cost stays bounded as history grows. Every existing family gives up at least one. Systems that consolidate records into extracted facts are compact, but discarded details are gone for good and consolidation can even introduce faulty memories. Keeping every record loses nothing yet must search an ever larger store. Parametric memories need training. Graph memories shatter an n-ary fact into disconnected triples, and file-system memories, with Claude Code's auto memory named explicitly, route retrieval through one-line summaries, so a detail no summary mentions is reachable only by string search. User-centric and agent-centric memory also face reversed challenges, broad and connected versus precise and refined, and MIRIX-2.0 builds one store for both.

Method

Storage and indexing are separate layers, and each layer owns one requirement.

Results

SettingMetricResultBaseline
LoCoMo (1,540 questions)Accuracy91.4%HyperMem 91.0% (authors' run; self-reported 92.7)
LongMemEval-SAccuracy72.6%Strongest baseline 58.3, +14.3
SHDocQA (AutoDream off)Accuracy83.0%HyperMem 52.0, +31.0
MHDocQA (AutoDream off)Accuracy71.0%HyperMem 49.0, +22.0
DetectiveQAAccuracy64.8%Below MemoryBank 67.6%
ALFWorld (GPT-6 Luna)Success76.4%AWM 73.6%, no memory 69.4%
WebShopSuccess37.4%AWM 36.0%, no memory 31.9%
MetaClaw-Bench (GPT-5.2)Official score49.2%MetaClaw skill library 44.0%

The ablation chain is clean: on LoCoMo, flat memory scores 85.7%, the hypergraph lifts it to 88.4%, AutoDream to 91.4%; on LongMemEval-S the three steps read 58.8%, 70.0%, and 72.6%, with the graph alone worth 11.2 points. Where plain vector retrieval struggles most, organizing evidence pays most. With AutoDream disabled, the system answers faster than HyperMem on all five timed benchmarks, 5.45 versus 10.55 seconds per question on SHDocQA, and on WebShop it spends the fewest training and test tokens among memory methods. The bill lands on ingestion: documents take 3.8 to 4.0 times longer than flat memory, and its ALFWorld test tokens (3.813M) are the highest among memory methods, because a retrieved procedure cuts interaction steps while thickening the context at each step.

Why it matters

The contribution is the constraint set, not any single score. Records are never rewritten, the graph is only an index, and consolidation changes how evidence is connected rather than what exists. That turns memory maintenance into an auditable operation where every answer lands on specific stored rows, worth more in production than another leaderboard point. Bounded evidence sets keep latency flat as history grows, exactly where agent products hurt. One store serves both personalization and agent skill reuse, which the paper claims as a first, giving later work a setup to compare against directly. No fine-tuning, closed and open models both work, and the code is public. The DetectiveQA loss to MemoryBank marks the boundary: organizing evidence well does not buy the inference a detective question needs.

Limitations

The authors are unusually direct. Ablations add components in groups, so the contributions of n-ary frames, cited rows, occurrence identity, and answer-time verification remain interpretations, not isolated measurements. User-memory results come from single runs, partly scored by an LLM judge. Agent results average three runs with large spreads, ALFWorld standard deviations of 6.2 and 14.0 points, so the mean edge over AWM is not stable. The GPT-5.2 MetaClaw baseline is the original paper's reported number, not a controlled rerun. Beyond that: the full system with AutoDream was never timed, so the latency claims cover only the graph configuration; the 0.4-point LoCoMo edge over HyperMem comes from the authors' own run of that baseline, and a single run cannot support it; and ingestion roughly 4x slower than flat memory is a real cost for write-heavy deployments.

Terms

Source

Related papers

All paper explainers