AuthorsXun Jiang, Feng Li, Han Zhao et al.
TL;DR
Long Term Memory + OMNE uses multi-agent LTM construction and utilization to enable AI self-evolution, achieving first place on the GAIA benchmark.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Long Term Memory notes that current foundation models compress vast data into an average model, which struggles with long-tail individual data and rare scenarios.
This limitation prevents OMNE from expressing personalized information, hindering AI self-evolution and multi-agent collaboration in complex, dynamic environments.
HOW IT WORKS
Long Term Memory introduces a Data Framework for LTM, Development Framework for LTM, and multi-agent collaboration mechanism to store and use refined individual interaction data.
An analogy is a human brain with cortical columns: each agent builds its own world model, while a shared long-term memory acts like a collective hippocampus and card catalog.
This KEY_MECHANISM lets Long Term Memory and OMNE represent sparse, evolving personal experiences beyond a fixed context window, enabling continual adaptation without retraining all parameters.
DIAGRAM
This diagram shows how Long Term Memory drives AI self-evolution across three phases of model evolution.
DIAGRAM
This diagram shows how Long Term Memory collects, structures, and deploys data in office collaboration and health management systems.
PROCESS
Data collection framework
Long Term Memory uses the Data Framework for LTM to gather fragmented digital and physical interaction data in office collaboration and health management.
Data synthesis techniques
Long Term Memory applies mixed synthetic strategies like role playing, EnvGen, AGA, and RTG synthesis to enrich sparse individual data.
Construction strategies of LTM
Long Term Memory transforms raw data into LTM via text summarization, data structuring, graph representation, vectorization, and model parameterization.
Development Framework for LTM
Long Term Memory powers OMNE, where each agent builds an independent world model over LTM and collaborates to adapt plans and decisions in real time.
KEY CONTRIBUTIONS
Definitions of AI Self-Evolution and LTM
Long Term Memory formalizes AI self-evolution and LTM, linking Differentiated Personalized Models and Long-Term Memory and Learning Ability as the core of personalized agents.
Data Framework for LTM
Long Term Memory proposes a data collection, analysis, and synthesis framework deployed in office collaboration and health management, including the world’s largest real-user voice dataset for mental health.
Development Framework for LTM
Long Term Memory introduces the multi-agent OMNE framework, where each agent maintains an independent world model over LTM, achieving first place on the GAIA benchmark.
RESULTS
GAIA benchmark rank
1st place
Top position over all reported GAIA baselines
Voice dataset scale
world’s largest
Largest real user mental health voice dataset reported by Long Term Memory
Business deployments
2 scenarios
Office collaboration and health management systems
Agent framework
multi agent
OMNE agents all equipped with independent LTM world models
On the GAIA benchmark, which tests complex real-world problem solving, Long Term Memory with OMNE reaches first place, showing that LTM-driven multi-agent personalization can handle diverse tasks. The large real-user mental health voice dataset and dual business deployments demonstrate that Long Term Memory scales beyond benchmarks into practical environments.
BENCHMARK
Relative GAIA benchmark ranking for OMNE vs generic agent baselines (higher rank is better).
KEY INSIGHT
Long Term Memory shows that self-evolution can emerge from limited, refined interaction data rather than ever larger pre-training corpora.
This challenges the common belief that scaling datasets alone is the main path to stronger AI, highlighting architecture and LTM as equally critical levers.
WHY IT MATTERS
Long Term Memory unlocks AI systems where thousands of OMNE agents evolve distinct world models over shared LTM while still sharing a core architecture.
Builders can now design ecosystems of personalized, collaborating agents that adapt over months of interaction, instead of repeatedly retraining monolithic foundation models.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…