Meta FAIR proposes Routed CL for agent memory routing
Meta FAIR proposed Routed Continual Learning (Routed CL), an architecture that groups historical trajectories into clusters and maintains a separate memory system for each one. Before a run, an LLM router chooses an existing cluster or creates a new one, while a Context Memory Adapter loads and updates the selected memory; the team reported improved normalized reward and 65% to 72% lower estimated API costs versus Global ICL in its evaluations. The study used 301 instances from six CL-Bench tasks and 97 banking instances from τ³-Bench across four models, but did not evaluate concurrent updates, stale-state reads, or cross-model cluster reuse.
Routed CL offers a task-aware way to route agent experiences into suitable memory clusters and representations without requiring agents to collaborate, as reported by 36氪.
面对相似任务时,让 agent 复用彼此积累的经验,是一种自然的思路。遗憾的是,并不存在一种在所有场景下都好用的经验保留方式,也无法提前知道某段经验会对哪些未来问题有帮助。
对于持续运行的 agent 系统而言,过往轨迹确实能减少后续任务的重复试错,可一旦不同任务的经验混在一起,也可能造成干扰;如果把所有历史轨迹都统一保留下来,细节确实最完整,却又会增加上下文和 API 成本;如果把所有经验都压缩成同一种记忆形式,又可能丢掉只对特定任务有用的信息。
为此,Meta FAIR 团队提出了“路由式持续学习”(Routed Continual Learning,Routed CL)。该架构会把历史轨迹归入不同的簇(cluster),并为每个簇单独维护一套记忆。每次运行前,基于 LLM 的路由器会依据当前问题、可用记忆系统的说明和已有簇的摘要,先判断这个任务和哪一类历史经验最接近,最后决定是复用某个簇,还是创建一个新簇。
#Meta FAIR#Routed Continual Learning#Context Memory Adapter#Gemini 3.7 Flash#GPT-5.6 Sol#Opus 4.7#Opus 5