MIRIX-2.0: Multi-Agent Memory System
AuthorsWei-Chieh Huang, Weizhi Zhang, Jiayan Nan, Peng Xia, Yu Wang
AffiliationsMIRIX AI
MIRIX-2.0 gives LLM agents a structured, provenance-tracked memory system that aims to improve long-term recall and reuse of skills.
Key results
Overall conversational question-answering accuracy.
Conversational memory accuracy; 14.3 points above the strongest baseline.
Document QA result with AutoDream disabled.
Multi-hop document QA result with AutoDream disabled.
Mean success with GPT-6 Luna.
Result with GPT-5.2, compared with 44.0% for MetaClaw’s skill library.
What the paper found
MIRIX-2.0 tackles long-term agent memory by separating complete, typed records from the graph used to find them. Six specialized memory agents write records that remain the system of record; a provenance hypergraph organizes episodic and semantic facts as n-ary frames, each linked to its source rows. At query time, a planner applies explicit retrieval budgets, keeping the evidence inspected bounded as history grows. Unlike summary-index approaches such as Anthropic’s Claude Code memory, the graph is an index rather than a replacement for the underlying records. AutoDream runs outside the write path: it merges duplicate entity anchors without rewriting evidence and distills task sessions into reusable skills. The system requires no fine-tuning and was evaluated with models including GPT-4.1-mini and GPT-4o-mini. It reaches 91.4% on LoCoMo and 72.6% on LongMemEval-S, 14.3 points above the strongest LongMemEval-S baseline. With AutoDream disabled, it scores 83.0% on SHDocQA and 71.0% on MHDocQA. Its procedural memory reaches 76.4% on ALFWorld with GPT-6 Luna and, with GPT-5.2, 49.2% on MetaClaw-Bench versus 44.0% for MetaClaw’s skill library. A trade-off is that graph construction makes document ingestion about four times slower than flat memory, although retrieval is faster than HyperMem on all five timed benchmarks.
Original abstract
Long-term memory lets LLM agents personalize, stay consistent, and reuse what worked before. Parametric memories need training, consolidation can lose details for good, and keeping every record leaves a growing store to search. We present MIRIX-2.0, a multi-agent memory system designed to be practical, accurate, and efficient at once. 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, and 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 yields the highest mean success among the compared memory methods on ALFWorld and WebShop and reaches 49.2% on MetaClaw-Bench, against 44.0% for MetaClaw's skill library. Six typed memory agents keep complete records as the system of record, and a provenance hypergraph indexes episodic and semantic records with n-ary frames that cite their rows. Retrieval runs under explicit budgets, so the evidence a question inspects does not grow with the history. AutoDream merges duplicate anchors outside the write path without rewriting any record, and it distills task sessions into skills. With AutoDream disabled, MIRIX-2.0 answers faster than HyperMem on all five timed benchmarks. It needs no fine-tuning, runs with closed and open models, and its code is public.
Read the original paperMore in AI Agents
Browse all 77 papers →From Attack Surface to Defense Architecture: Compositional Security for Agentic Coding Harnesses
Simon Hegelich, Bardh Prenkaj, Habiba Sarhan, Morteza Shahrezaye, Kolja Hegelich, Christian Montag, Ralf Schneider, Erik Cambria
A security architecture aims to keep coding agents useful while preventing malicious instructions from turning their tools into an attack platform.
HarnessFlow: MCTS-Guided Harness Optimization for LLM Agents
Zhi Hong, Qianlong Wang, Hongchen Wei, Aidan Wu, Daohong Ying, Mingze Kong, Bei Liu, Chong Luo
HarnessFlow uses tree search and a coding agent to automatically refine the software harnesses that guide LLM agents.
Test-Time Compute Scales Better with Parallel Agent Teams
Junlin Wang, Haozhe Chen, Ang Li, Rahul Chalamala, Zhuang Liu
Instead of choosing between many solo agents and one huge team, run many small teams in parallel and keep the strongest result.