当前AI编码代理的记忆存储系统正面临严重的碎片化问题,缺乏必要的互操作性。1 现有生态中主要存在四种互不兼容的记忆格式,包括AGENTS.md或.cursor规则、Anthropic内存工具、GitHub Copilot事实库以及JSONL知识图谱。1 这种混乱局面被类比为可观测性领域在2019年OpenTelemetry统一标准之前的状态,当时OpenTracing和OpenCensus通过合并才确立了统一的数据模型、语义约定及OTLP协议。1
为解决这一问题,开发者社区提议参照OpenTelemetry模式,建立一套独立于特定厂商的Agent记忆数据标准。1 该提议的核心要素涵盖地址级别(如global、project、repo等)、身份ID、生命周期管理(包含过期时间)、写入者类型(区分人类与模型)以及内容类型。1 出于安全考量,新标准要求记录记忆的写入者以防范提示注入攻击,并支持签名验证机制。1
目前已有多个项目涉足相关探索,包括MacPaw Research推出的Portable Memory、计划于2026年4月发布草案的Open Memory Protocol以及Cognee开发的COGX格式。1 文章作者详细对比了Copilot、LangGraph、Graphiti及LoreKit等现有方案的特点与不足,指出标准应仅规定数据规范,而将存储后端、排序算法和用户界面留给各工具进行竞争创新。1
The current landscape of AI coding agent memory storage is characterized by extreme fragmentation and a lack of interoperability, mirroring the state of the observability sector prior to its unification in 2019 1. This disjointed environment prevents different tools from sharing context effectively within the AI agent ecosystem 1. The situation draws a direct parallel to the period before OpenTracing and OpenCensus merged to form OpenTelemetry, which established a unified data model, semantic conventions, and the OTLP protocol 1.
To address these deficiencies, developers are calling for the community to adopt an OpenTelemetry-like approach for agent memory standards. Key proposals include a shared specification draft based on YAML frontmatter that defines essential elements such as address levels (global, project, repo), unique identity IDs, lifecycle expiration times, writer types (human or model), and content formats 1. Security considerations are central to this proposal, with requirements to track who wrote specific memories to mitigate prompt injection attacks and support signature verification 1. Notably, the proposed standard explicitly excludes specifications for storage backends, sorting algorithms, or user interfaces, allowing vendors to compete on those implementation details while maintaining a common data exchange layer 1.
Several existing solutions illustrate the current incompatibility issues, including AGENTS.md/.cursor rules, Anthropic’s memory tools, GitHub Copilot’s fact library, and JSONL knowledge graphs 1. These distinct formats cannot easily communicate with one another, highlighting the urgent need for a vendor-neutral standard 1. Related initiatives mentioned in discussions around this topic include MacPaw Research’s Portable Memory, the Open Memory Protocol (a draft dated April 2026), and Cognee’s COGX format 1. By establishing a common data model, the industry aims to break down silos and enable more robust, interconnected AI agent workflows without locking users into specific proprietary ecosystems 1.
评论
还没有评论,欢迎留下第一条。