MIRIX-2.0 keeps complete agent memory records and retrieves them under explicit budgets
The MIRIX-2.0 paper splits long-term agent memory across six typed agents with a provenance hypergraph, reaching 91.4% on LoCoMo and 72.6% on LongMemEval-S without fine-tuning.
MIRIX-2.0 is a multi-agent memory system for LLM agents, described in a paper posted to arXiv on 2026-11-09 as arXiv 2611.12068 and listed on Hugging Face daily papers. The paper has five authors, and the code is public in the Mirix-AI/MIRIX repository on GitHub.
The starting point is a trade-off in how agents remember. Long-term memory lets an agent personalize its answers, stay consistent and reuse what worked before, but according to the abstract each common approach has a cost: parametric memories need training, consolidation can lose details for good, and keeping every record leaves a growing store to search. MIRIX-2.0 is presented as a system designed to be practical, accurate and efficient at once.
How the memory is organized
Six typed memory agents keep complete records, and those records act as the system of record rather than a compressed summary. On top of them, a provenance hypergraph indexes episodic and semantic records with n-ary frames, and each frame cites the rows it was built from. That design means an answer can be traced back to the stored records behind it.
Retrieval runs under explicit budgets. The abstract states that the evidence a question inspects does not grow with the history, which targets the search cost of keeping every record without throwing records away.
A component called AutoDream handles maintenance outside the write path. It merges duplicate anchors without rewriting any record, and it distills task sessions into skills, which is how the same system also serves as procedural memory.
Reported results
The paper compares MIRIX-2.0 with other memory systems on question answering and on agent tasks:
- It is the most accurate of the compared systems on four of five QA benchmarks.
- It reaches 91.4% on LoCoMo and 72.6% on LongMemEval-S, where it leads the strongest baseline by 14.3 points.
- With AutoDream disabled, it gains 31.0 and 22.0 points over HyperMem on single-hop and multi-hop document QA.
- As procedural memory, it has the highest mean success among the compared memory methods on ALFWorld and WebShop.
- On MetaClaw-Bench it reaches 49.2%, against 44.0% for the MetaClaw skill library.
Speed is reported as well. With AutoDream disabled, MIRIX-2.0 answers faster than HyperMem on all five timed benchmarks.
Practical requirements
The system needs no fine-tuning and runs with both closed and open models, so it is meant to sit next to an existing agent rather than require a separately trained memory model. The abstract does not list the specific models, hardware or cost figures behind the experiments, so those details have to come from the full paper and the repository.
Source: MIRIX-2.0: Multi-Agent Memory System (arXiv 2611.12068)