Source-linked AI summary
Rethinking Memory as Continuously Evolving Connectivity
Jizhan Fang, Buqiang Xu, Zhixian Wang, Haoliang Cao, Xinle Deng, Baohua Dong, Hangcheng Zhu, Ruohui Huang, Gang Yu, Ying Wei, Guozhou Zheng, Feiyu Xiong, Haofen Wang, Huajun Chen, Ningyu Zhang
TL;DR
Static memory repositories and fixed retrieval pipelines struggle to adapt connectivity to changing agentic tasks. FluxMem models memory as an editable heterogeneous graph, refines it online through feedback, and consolidates successful trajectories offline. Across LoCoMo, Mind2Web, and GAIA, it achieves state-of-the-art performance, while its evaluation does not systematically measure deployment overhead or open-world distribution shifts.
Problem
Static memory systems use fixed representations and retrieval pipelines that cannot dynamically refine memory structures from environmental feedback across diverse tasks.
Method
FluxMem models memory as a dynamically editable heterogeneous graph and evolves it through initial connection formation, feedback-driven refinement, and long-term consolidation.
Results
FluxMem achieves state-of-the-art performance across LoCoMo, Mind2Web, and GAIA, including 95.06 average accuracy on LoCoMo.
Takeaways & Limitations
FluxMem provides a connectivity-evolving memory framework for adaptation and generalization in complex agentic environments.
Takeaways & Limitations
The evaluation does not systematically measure latency, API cost, or token consumption and uses static benchmarks that do not fully simulate open-world distribution shifts.
Abstract
from arXiv · showhide
Existing memory-augmented LLM agents often treat memory as a static repository with pre-defined representations and fixed retrieval pipelines, which is brittle in dynamic agentic environments where feedback, task variation, and heterogeneous signals continuously reshape what should be remembered and how it should be connected. To address this, we propose FluxMem, a connectivity-evolving memory framework that models memory as a heterogeneous graph and progressively refines its topology through three stages: initial connection formation, feedback-driven refinement, and long-term consolidation. During execution, FluxMem repairs missing links, prunes interference, aligns abstraction granularity, and distills recurrent successful trajectories into reusable procedural circuits, guided by one metric for memory generalizability and evolutionary maturity. Across three fundamentally distinct benchmarks including LoCoMo, Mind2Web, and GAIA, FluxMem achieves consistent state-of-the-art performance, demonstrating strong adaptation and generalization in complex agentic environments. The code will be open-sourced in https://github.com/zjunlp/LightMem.
1 Introduction
FluxMem reframes memory as continuously evolving connectivity because static systems struggle to adapt links, units, and consolidation to changing tasks. It introduces a three-stage evolutionary pipeline and reports state-of-the-art results across three benchmarks.
- Memory supports long-horizon agents by distilling facts, experiences, and skills from past interactions and retrieving them for related tasks.
- Useful memory depends on accessing the right memories at each decision step, motivating memory connectivity as the central problem.
- Challenges: Static, hand-crafted pipelines cannot reliably refine memory structures from environmental feedback, causing inaccurate connections and inflexible units.
- Challenges: Existing systems preserve trajectories as isolated instances, so agents repeatedly reconstruct associations instead of consolidating enduring structural patterns.
- Method: FluxMem models memory as a dynamically editable heterogeneous graph and evolves it through connection formation, feedback-driven refinement, and long-term consolidation.
- Results: 95.06 average accuracy on LoCoMo, 8.1 Cross-Task success rate on Mind2Web, and 64.85 average success rate on GAIA demonstrate consistent state-of-the-art performance.
2 FluxMem Memory Architecture
FluxMem represents memory as a heterogeneous graph with semantic, episodic, and procedural layers. Task-specific activated subgraphs connect evidence, experiences, and reusable skills into evolving contexts.
- FluxMem models memory as a heterogeneous graph G = (V, E) spanning semantic, episodic, and procedural layers.
- Semantic nodes store factual knowledge, episodic nodes record state-action trajectories, and procedural nodes encode distilled reasoning templates.
- The episodic layer orchestrates interactions between static knowledge and distilled skills.
- Stages I and II operate online at step-wise granularity, whereas Stage III operates offline for performance optimization and memory consolidation.
- Grounding edges connect semantic facts to task steps, while distillation edges connect past episodes to reusable procedural skills.
- At each step, FluxMem selects a task-specific local subgraph whose activated nodes and connections determine the working context.
3 Three-Stage Memory Evolution
FluxMem evolves task-local memory through online retrieval and feedback-based editing, then consolidates successful episodes offline into validated procedural skills. The stages progressively turn tentative links into task-optimized circuits.
- Stage I: Initial Connection Formation: Stage I retrieves semantic facts, similar episodes, and inherited procedural skills to form an initial step-local subgraph.
- Stage I: Initial Connection Formation: Semantic retrieval combines dense similarity, sparse lexical matching, and LLM verification to connect observations with supporting factual knowledge.
- Stage II: Feedback-Driven Connectivity Refinement: Stage II uses execution feedback to expand missing links, prune interference, and reshape memory units when their abstraction granularity mismatches the task.
- Stage II: Feedback-Driven Connectivity Refinement: Refinement updates the local subgraph and terminates when execution succeeds or the predefined refinement-round limit T is reached.
- Stage III: Long-Term Connection Consolidation: Stage III clusters completed trajectories by semantic similarity and induces procedural skills capturing patterns shared across episodes.
- Stage III: Long-Term Connection Consolidation: PEMS-guided consolidation re-runs source episodes, scores each skill using success, length, and embedding change, and rewrites low-scoring skills until improvement falls below ε.
4 Experiments
FluxMem is evaluated across LoCoMo, Mind2Web, and GAIA, where it reports strong performance across long-context reasoning, web navigation, and general assistant tasks. Ablations and case studies examine which evolutionary stages and feedback-driven edits matter across task types.
- Main Results: FluxMem achieves state-of-the-art performance across LoCoMo, Mind2Web, and GAIA, covering long-context reasoning, web navigation, and general assistant tasks.The evaluation includes multiple backbones and realistic web-navigation settings.
- Main Results: 95.06 average LMJ with GPT-4.1-mini surpasses Full Context at 81.23 and EverMemOS at 93.05 on LoCoMo.With Qwen3-30B-A3B-2507-Instruct, FluxMem reaches 93.44 versus 74.87 for Full Context.
- Main Results: 8.1 Cross-Task Success Rate with GPT-4.1-mini exceeds AWM at 3.6 in realistic Mind2Web, while Gemini-2.5-flash reaches 9.6 versus 5.6.FluxMem also yields the highest SSR and AF1 across sub-categories and backbones.
- Main Results: 12.73% absolute improvement raises GAIA average success from 52.12 to 64.85 with Kimi K2.With GPT-5-mini, FluxMem reaches 53.85 success rate on Level 3 tasks.
- Ablation Study: Removing Stage II lowers LoCoMo average LMJ from 95.06% to 85.32% with GPT-4.1-mini and from 93.44% to 84.74% with Qwen3-30B-A3B.The reported effect is larger than those of the other two stage ablations.
- Ablation Study: Removing Stage III lowers a Mind2Web first-subcategory success rate from 8.1% to 3.2% with GPT-4.1-mini.The authors identify long-term consolidation as the primary performance driver for complex, multi-step web navigation.
- Ablation Study: Increasing Stage II refinement rounds from T = 0 to T = 5 raises LoCoMo average performance from 85.32% to 95.06%.The improvement is reported as consistent and monotonic across sub-categories and the overall score.
- Case Study: In the GAIA case study, feedback prunes an ineffective visualization connection, adds a Python analysis API, and reshapes an overly coarse skill.The edits follow an execution failure and self-verification of the inherited procedure.
5 Related Work
Related work includes hierarchical structured memory systems and self-evolving agent memory. These approaches organize memory through topologies or evolve contextual repositories from historical trajectories.
- Hierarchical Structured Memory Systems: Hierarchical memory systems include tree-based, graph-based, pyramid, and heterogeneous multi-layer structures with different organization and querying strategies.The passage contrasts hierarchical abstraction with graph connectivity dynamics and coarse-to-fine querying.
- Self Evolving Agent Memory: Self-evolving agent memory research includes contextual repositories that distill experiences from historical trajectories to enhance agent capabilities.The cited paradigm includes Expel, AWM, and ReasoningBank.
6 Conclusion
FluxMem conceptualizes agent memory as dynamic connectivity and enables autonomous adaptation through a three-phase evolutionary process. It achieves state-of-the-art results across LoCoMo, Mind2Web, and GAIA.
- FluxMem models agent memory as dynamic connectivity rather than a static repository.
- A three-phase evolution enables autonomous memory adaptation.
- FluxMem achieves state-of-the-art results across LoCoMo, Mind2Web, and GAIA.
Limitations
The experimental design has limitations involving computational overhead, benchmark protocols, and hyperparameter sensitivity. These constraints matter particularly for resource-constrained, open-world, and heterogeneous deployments.
- Computational Overhead of Closed-Loop Operations: Stages II and III use iterative LLM calls, but latency, API cost, and token consumption are not systematically measured.
- Static Benchmark Protocols: Static LoCoMo, Mind2Web, and GAIA datasets do not fully simulate continuous open-world distribution shifts or streaming environments.
- Hyperparameter Sensitivity: The framework introduces control thresholds, yet lacks comprehensive sensitivity analysis across backbones and heterogeneous domains.
Use of AI Assistants
AI language models were used only for linguistic refinement, grammar correction, and stylistic polishing. The human authors independently conceived, developed, and verified the core research and retained responsibility for its accuracy and originality.
- AI language models were used exclusively for linguistic refinement, grammar correction, and stylistic polishing.
- Human authors independently conceived, developed, and verified the research concepts, experiments, methods, narrative, and figures.
- The authors retain responsibility for the work’s scientific accuracy, originality, and integrity.
A.1 Detailed Dataset Statistics
The evaluation covers long-context reasoning, web navigation, and generalist assistant tasks using detailed benchmark subsets. These datasets span varying task complexity and test generalization across multiple settings.
- LoCoMo: LoCoMo contains 10 extensive conversations averaging 588 turns and 16,618 tokens, with 1,540 human-annotated questions.The questions include 841 single-hop, 282 multi-hop, 321 temporal-reasoning, and 96 open-domain items.
- Mind2Web: Mind2Web contains 2,350 open-ended tasks from 137 websites across 31 domains, averaging 1,135 DOM elements and 7.3 actions per task.
- Mind2Web: Mind2Web evaluates generalization across cross-task, cross-website, and cross-domain scenarios.
- GAIA: GAIA includes 165 tasks across three difficulty levels, from basic tool use and retrieval to long-horizon multimodal execution.The set contains 53 Level-1, 86 Level-2, and 26 Level-3 tasks.
A.2 Implementation Details
The implementation details specify benchmark-specific retrieval, evaluation, and agent configurations across LoCoMo, Mind2Web, and GAIA. Baselines generally follow their native pipelines, while FluxMem is evaluated within the Flash-Searcher framework on GAIA.
- LoCoMo: LoCoMo retrieval baselines use text-embedding-3-small, with Zep and Mem0 retrieving top-10 memories per speaker and A-MEM globally retrieving top-40 entries.MemoryOS instead uses a three-tier hierarchy with exhaustive STM recall and selective MTM retrieval.
- LoCoMo: MemoryOS uses exhaustive recall for Short-Term Memory and two-stage selection for Mid-Term Memory, including top-5 segments and top-10 dialogue pages.
- Mind2Web: Mind2Web baselines are evaluated offline through their native pipelines with GPT-4.1-mini and Gemini-2.5-flash as base models.AWM's original setting filters the current page to top-k = 5 candidates, which can discard substantial information because pages contain many elements.
- GAIA: GAIA evaluates every method with a single run per task, spanning diverse model, knowledge-base, orchestration, and agent-architecture configurations.The listed systems include Langfun, OpenAI Deep Research, Magnetic-1, Agent KB, Alita, and Smolagents.
- GAIA: Flash-Searcher, MemEvolve, and FluxMem are evaluated within the Flash-Searcher framework, configured with 40 maximum steps.