FluxMem: Rethinking AI Agent Memory as Continuously Evolving Connectivity

> arXiv: 2605.28773 | Rethinking Memory as Continuously Evolving Connectivity > Authors: 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 > Published: 2026-05-27

Introduction: The AI That Forgets at Critical Moments

The post opens with a relatable scenario: a personal assistant who knows your preferences (oat lattes, no cilantro, gym on Wednesdays) but still recommends Sichuan food in a new city—forgetting that spicy food upsets your stomach. Current AI agents face the same problem: their memory behaves like a static database rather than a living, self-reorganizing network.

Why Static Memory Fails in Dynamic Worlds

Most memory-augmented LLM agents resemble a giant library: each memory is a book retrieved by keyword similarity. The post identifies three fatal flaws:

1. Entries rarely change once stored — no rewriting or new connections form over time. 2. Entries are isolated — related memories lack genuine links; the agent must compare them manually. 3. Retrieval is fixed — the same strategy on day 100 as on day 1.

In dynamic agentic environments, feedback, task changes, and heterogeneous signals constantly reshape *what* should be remembered and *how* it should be connected. FluxMem's core insight: memory should not be a warehouse but an ever-evolving network.

The FluxMem Architecture

Heterogeneous Memory Graph

FluxMem models memory as a heterogeneous graph, distinguishing memory types (facts, procedures, context, preferences/emotional signals) with different node attributes and edge semantics (e.g., causal edges vs. temporal edges).

Three-Stage Evolution

1. Initial Connection Formation — new information is linked via semantic similarity and co-occurrence patterns, forming weak initial connections. 2. Feedback-Driven Refinement — task feedback drives three operations:

• Repairing missing links (e.g., realizing a missing precondition in a failed task)
• Pruning interfering connections (weakening—not deleting—contradicted memories by reducing connection weights)
• Aligning abstraction granularity (lifting lessons from specific cases to general principles)
3. Long-Term Consolidation — repeatedly verified successful trajectories are distilled into procedural circuits—directly executable "muscle memory" that bypasses retrieval and reasoning, mirroring the shift from declarative to procedural memory.

A Unified Metric

The paper proposes memory generalizability and evolutionary maturity, measuring (1) whether memories transfer from one context to novel situations, and (2) how many feedback-refinement cycles a connection has survived.

Experimental Validation

FluxMem was evaluated on three distinct benchmarks:

• LoCoMo (long multi-turn conversation): graph connectivity keeps early information reachable via multiple paths, avoiding recency-biased retrieval failure.
• Mind2Web (real-world web tasks): feedback-driven refinement lets the agent re-prioritize memories after each page transition.
• GAIA (general AI assistance): cross-task generalization—search strategies learned in web tasks transfer to QA; preference tracking transfers to recommendations.

Key result: FluxMem achieves state-of-the-art performance on all three benchmarks, indicating a general-purpose memory framework rather than a dataset-specific optimization.

Philosophical and Neuroscientific Framing

The post argues that knowledge is not accumulated information but *patterns of connection*, and that dynamic environments demand memory systems that evolve along with adaptation rather than being used after adaptation completes. FluxMem engineering-analogizes several brain mechanisms:

• Connection strengthening/weakening → long-term potentiation (LTP) and depression (LTD)
• Procedural circuit distillation → hippocampal-to-cortical memory consolidation
• Interference pruning → active forgetting

Conclusion

FluxMem envisions agent memory that grows, repairs, and evolves on its own—no longer a burden requiring manual curation. The paper states code will be open-sourced at https://github.com/zjunlp/LightMem.

Reference: Fang, J., Xu, B., Wang, Z., et al. (2026). *Rethinking Memory as Continuously Evolving Connectivity*. arXiv preprint arXiv:2605.28773.