做智能体记忆系统的开发者终于有了一个跳出传统检索范式的方案——FluxMem 用图结构动态修复连接，实测效果显著，值得深入研究其实现细节。
论文提出 FluxMem 记忆系统，将智能体记忆视为不断变化的连接网络，而非静态文件柜。它存储事实、任务片段和可复用技能作为图中的连接点，在任务执行时先收集有用记忆，再根据反馈修复连接（增删链接或调整细节）。长期运行中，重复成功的任务路径会自动转化为可复用技能。在长对话记忆、网页导航和通用助手任务上，FluxMem 取得显著提升，包括 LoCoMo 上 95.06% 平均准确率和 GAIA 上 12.73 个百分点的增益。该研究将智能体记忆从“存储-检索”范式转向“持续修复和强化有效连接”。
AI agents should treat memory as a changing web of…
AI agents should treat memory as a changing web of useful connections, not static storage.
Most agent memory systems retrieve old facts as if the past were a filing cabinet.
The paper proposes FluxMem, a memory system that stores facts, past task episodes, and reusable skills as connected pieces in a graph.
When the agent works on a task, FluxMem first gathers likely useful memories, then uses feedback from the task to fix the memory connections by adding missing links, removing bad ones, or rewriting memories at the right level of detail.
Over time, it also turns repeated successful task paths into reusable skills, so the agent does not need to rebuild the same reasoning pattern again and again.
The authors tested FluxMem on long conversation memory, web navigation, and general assistant tasks, which checks whether the idea works across very different agent problems.
FluxMem got stronger results than the compared memory systems, including 95.06 average accuracy on LoCoMo and a 12.73-point gain on GAIA with Kimi K2.
The big deal is that the paper shifts agent memory from “store and retrieve” toward “keep repairing and strengthening the connections that actually help the agent act.”
----
Link – arxiv. org/abs/2605.28773
Title: "Rethinking Memory as Continuously Evolving Connectivity"