MIRIX-2.0: Multi-Agent Memory System

Why long-term memory is still difficult for AI agents

Long-term memory is essential if an AI agent must personalize its behavior, remain consistent across sessions, or improve by reusing successful strategies. Yet memory creates a difficult engineering trade-off. Compressing conversations into summaries or extracted facts can discard names, dates, quantities, and distinctions between similar events. Keeping every record preserves detail, but retrieval becomes slower and noisier as the history grows. Parametric memory avoids some retrieval work, but it requires training and is not practical with closed models.

In MIRIX-2.0: Multi-Agent Memory System, Wei-Chieh Huang, Weizhi Zhang, Jiayan Nan, Peng Xia, and Yu Wang propose a layered design intended to address accuracy, traceability, and efficiency together. The system separates durable evidence from the structures used to find it, while supporting both user-centric memory and agent-centric procedural memory.

Complete records, structured indexing

MIRIX-2.0 uses six typed memory agents managed by a Meta Memory Manager: Core, Episodic, Semantic, Procedural, Resource, and Knowledge Vault. These agents write records to an evidence plane that acts as the system of record. The authors emphasize that records remain complete and are committed before graph construction begins, so a failed indexing operation does not erase the underlying memory.

Only episodic and semantic records are indexed by the system’s provenance hypergraph. Rather than replacing source records, the graph stores entities, concepts, events, relations, roles, and time information, with each fact pointing back to the rows that support it. The use of n-ary frames is important: a fact involving several participants can remain one connected structure instead of being split into disconnected binary relationships. Dates, prices, counts, and measurements are retained as literals attached to the relevant assertion.

The design also distinguishes semantic identity from event identity. Repeated evidence for a durable claim can accumulate around one semantic frame, while separate undated occurrences receive separate episodic identities. This is intended to prevent two similarly worded events from being incorrectly collapsed.

Bounded retrieval and offline consolidation

Retrieval is organized as a bounded evidence-construction process rather than an unrestricted similarity search. A deterministic planner classifies the query, selects relevant anchors, traverses a limited graph neighborhood, and retrieves only the cited source rows. Separate budgets control base-recall rows and relation-cited rows. The authors state that the inspected evidence therefore remains bounded by the query operator instead of growing with the entire history.

The answerer can make multiple retrieval calls, and a verification step can request another search when a draft contains unsupported numbers, missing values, inconsistent counts, or other detectable conflicts with the evidence ledger. This is a useful provenance pattern: the graph proposes relationships, but the answer is grounded in the original records that justify them.

MIRIX-2.0 also includes AutoDream, which runs outside the critical write path. For semantic memory, it compares nearby anchors of compatible types and asks a language model whether they refer to the same entity. Confirmed merges rewire graph connections but do not rewrite stored evidence, event frames, or source records. For procedural memory, AutoDream distills task sessions into reusable lessons and allows a curator to create or revise skills with citations to the experiences behind those edits.

What the evaluations show

The authors report strong results on conversational and document question answering. MIRIX-2.0 reaches 91.4% on LoCoMo and 72.6% on LongMemEval-S, leading the strongest reported comparison on the latter by 14.3 percentage points. On LoCoMo, the flat-memory configuration scores 85.7%, graph indexing raises this to 88.4%, and graph indexing plus AutoDream reaches 91.4%. On LongMemEval-S, the corresponding scores are 58.8%, 70.0%, and 72.6%.

With AutoDream disabled, the system reports 83.0% on single-hop document QA and 71.0% on multi-hop document QA, exceeding the cited strongest baseline by 31.0 and 22.0 points. However, DetectiveQA is an exception: MIRIX-2.0 scores 64.8%, below MemoryBank and HyperMem in the reported comparison. This suggests that structured evidence retrieval does not automatically solve more demanding reasoning tasks.

The procedural-memory results are also positive, though more mixed. On ALFWorld, MIRIX-2.0 records the highest mean success among the compared memory methods with both tested backbones: 76.4% with GPT-6 Luna and 61.7% with GPT-5.6 Luna. On WebShop, it reaches 37.4%, compared with 36.0% for AWM and 31.9% without memory. On MetaClaw-Bench, the system reaches 49.2% with GPT-5.2 versus 44.0% for MetaClaw’s skill library, while the advantage is much smaller with the other tested backbone.

The efficiency trade-off

The results do not support a simple claim that the full architecture is always cheaper. With AutoDream disabled, MIRIX-2.0 answers faster than HyperMem on all five timed benchmarks and uses fewer answer tokens on four of them. But graph construction makes document ingestion roughly 3.8 to 4.0 times slower than flat memory. The full system with AutoDream enabled was not timed, so the reported latency results characterize the graph configuration without consolidation rather than the complete deployment.

Procedural memory can also increase token use even when it reduces action steps. On ALFWorld, MIRIX-2.0 uses the highest test-token total among the memory methods in the reported comparison. On WebShop, it uses the fewest training and test tokens among those methods while achieving the highest mean success. The practical lesson is that memory should be evaluated against task success, latency, and token cost together.

Limitations and practical implications

The paper’s ablations add major components in groups. They do not independently measure the contributions of n-ary frames, source-row citations, occurrence identities, answer-time verification, or each procedural-memory component. User-memory results include single-run comparisons and some LLM-based judging, while agent results average three runs and include substantial variation. Some baseline figures are reported by their original authors rather than reproduced under a fully controlled protocol.

For teams building long-running agents, MIRIX-2.0 offers a clear architectural principle: keep original evidence immutable, treat graphs as navigational indexes, make retrieval budgets explicit, and run consolidation away from the write path. That approach can improve auditability and reduce the risk that a faulty summary silently becomes the only version of the truth. It does not remove the need for privacy controls, deletion and correction workflows, access isolation, or human review of learned skills. The authors specifically warn that personal memories can contain sensitive information and that injected or incorrect content may persist.

MIRIX-2.0 is therefore best understood not as proof that one memory architecture solves long-horizon agents, but as a detailed demonstration of how provenance, bounded retrieval, and reversible consolidation can be combined in a deployable system.


Sources

AI-assisted research summary. Check the original paper for complete methods, results, and limitations.

Avatar photo

Tech News

Jeff is a passionate blog writer who shares clear, practical insights on technology, digital trends and AI industries. With a focus on simplicity and real-world experience, his writing helps readers understand complex topics in an accessible way. Through his blog, Jeff aims to inform, educate, and inspire curiosity, always valuing clarity, reliability, and continuous learning.

↑ ↓ navigate Enter open Esc close