Graph Memory vs Flat Retrieval
Finally, a good paper testing if graph memory actually beats flat retrieval for…
Agent开发者必读，这篇论文用严谨的基准测试证明了图记忆在长程任务中未必优于向量检索，并给出了具体的性能损耗场景与遗忘优化数据。
Finally, a good paper testing if graph memory actually beats flat retrieval for long-term agents.
最后，一篇优秀的论文测试了图内存是否真的能在长期智能体中优于扁平检索。
(bookmark this one)
(收藏这篇)
Researchers extract each conversational turn into typed nodes and attributed edges, answer from a two-hop subgraph, and hold the candidate-generation budget fixed at five retrieval roots.
研究人员将每个对话轮次提取为带类型的节点和属性边，从两跳子图中进行回答，并将候选生成预算固定在五个检索根节点上。
On LongMemEval the graph gets token F1 0.42 against 0.47 for a flat vector baseline, and a paired bootstrap over 500 questions puts the gap at -0.050 (95% CI -0.085 to -0.016).
在 LongMemEval 上，图方法的 token F1 为 0.42，而扁平向量基线为 0.47；对 500 个问题进行的配对 bootstrap 检验显示差距为 -0.050（95% CI -0.085 至 -0.016）。
The damage concentrates on questions that require recalling a specific prior assistant turn, where judged correctness falls from 0.911 to 0.607. Splitting a turn into entities discards the surface form those questions depend on.
性能下降主要集中在需要回忆特定先前助手回复的问题上，其中判断正确率从 0.911 降至 0.607。将一轮对话拆分为实体丢弃了这些问题所依赖的表面形式。
The forgetting module fares much better. One pruning pass over a persistent 27,021-node graph, scored on recency, access frequency, degree centrality and age, removes 9.8% of nodes and 9.5% of stored bytes with token F1 unchanged.
遗忘模块的表现要好得多。对包含 27,021 个节点的持久化图进行一次剪枝，根据近期性、访问频率、度中心性和年龄进行评分，移除了 9.8% 的节点和 9.5% 的存储字节，同时 token F1 保持不变。
Paper: https://arxiv.org/abs/2608.28978
论文：https://arxiv.org/abs/2608.28978
Chat with Paper: https://academy.dair.ai/papers/selective-forgetting-a-graph-based-memory-framework-for-long-term-llm-agents-2608.28978
与论文对话：https://academy.dair.ai/papers/selective-forgetting-a-graph-based-memory-framework-for-long-term-llm-agents-2608.28978
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力