Agent memory needs its OpenTelemetry moment
191 points
• 4 days ago
• Article
Link
当前的 AI agent memory 格局高度碎片化，每个工具都使用各自专有的存储格式。无论是像 AGENTS.md 这样的指令文件，还是专门的 memory 目录，亦或是复杂的 knowledge graph，这些系统本质上都被彼此割裂。由于没有任何工具能够读取或理解另一个工具的输出，一个 agent 所汲取的宝贵经验就只能被困在原地，迫使用户不断重复搭建同样的上下文。这个问题并不在于工具不够多，而在于缺少对数据应如何存储与共享达成共识的共同协议。
这种局面与 2019 年之前的 observability 状态十分相似：当时各家供应商迫使开发者使用各不相同的 instrumentation library 和数据格式。 OpenTelemetry 的出现通过建立共享数据模型、 semantic conventions 以及统一协议来化解这一困境，使供应商能够在 analytics 和 visualization 上展开竞争，而不必把精力耗在数据采集本身。 Agent memory 现在也站在类似的十字路口。只要社区能就 agents 如何存储与交换 memories 制定一套共享规范，就有机会跨过当前的重复建设困境，形成稳健且可互操作的 ecosystem 。
当前格式的多样性背后，反映出一种共同的核心理念：大多数系统实际上只是在保存可供未来使用的笔记，而这些笔记往往只在有限时间内有效。尽管已有一些提案试图通过可导出的 bundles 来解决问题，但它们更着眼于在供应商之间搬运数据，而不是提供 live addressing 。 Live addressing 能让多个 agents（例如 CI pipeline 、编辑器助手和评审 bot）同时读取并写入同一份 memory space 。要实现这一点，统一规范需要为 identity 、 location 、 authorship 与生命周期定义标准字段，以便不同工具能解析并复用同一类记录。
一种潜在的共享规范可以采用把 envelope 与 content 分离的记录结构。使用类似 Markdown 、并配以结构化 frontmatter 的标准格式，开发者就能在其中加入关键元数据，例如 agent 的 identity 、 memory 的地址层级以及其过期日期，同时让实际的 memory content 保持足够灵活。通过对这些 envelope 字段进行标准化，并为 semantic conventions 建立共享注册表，行业就能在维持 tool-agnostic 的 memory store 的同时，让各家供应商继续在具体的存储与处理逻辑上展开竞争。
尽管有人怀疑另一个标准是否只是"再添噪音"，但开源发展的历史表明：当社区选择合并已有成果、而不是各自推出孤立的替代方案时，碎片化往往会被克服。关于 prompt injection 和安全性的担忧是有道理的，但这些风险可以通过记录每条 memory 来源信息的元数据来应对。围绕如何处理代码依赖、过期的默认策略以及地址层级等开放问题，仍需要协作攻关；不过，迈向统一的 memory spec 无疑是让 AI agent 能力走向成熟的下一步必经之路。
The current landscape of AI agent memory is highly fragmented, with every tool utilizing its own proprietary storage format. Whether it is an instruction file like an AGENTS.md, a dedicated memory directory, or a complex knowledge graph, these systems are effectively siloed. Because no tool can read or interpret the output of another, the valuable lessons learned by one agent remain trapped, forcing users to repeatedly rebuild the same context. This issue is not due to a lack of tools, but rather a lack of a common agreement on how data should be stored and shared.
This situation mirrors the state of observability before 2019, when various vendors forced developers to use disparate instrumentation libraries and data formats. The emergence of OpenTelemetry solved this by establishing a shared data model, semantic conventions, and a unified protocol, allowing vendors to compete on analytics and visualization rather than data capture. Agent memory is now at a similar crossroads. By establishing a shared specification for how agents store and exchange memories, the community could move past the current state of duplication and create a robust, interoperable ecosystem.
The current variety of formats reveals a shared core philosophy, as most systems essentially store notes for future use that are true for a limited duration. While some existing proposals have attempted to address this through exportable bundles, they primarily focus on moving data between vendors rather than enabling live addressing. Live addressing would allow multiple agents, such as a CI pipeline, an editor assistant, and a review bot, to read and write to the same memory space simultaneously. To achieve this, a unified spec would need to define standard fields for identity, location, authorship, and lifecycle, ensuring that different tools can parse and utilize the same records.
A potential shared specification could utilize a record structure that separates the envelope from the content. By using a standard format like Markdown with structured frontmatter, developers could include essential metadata, such as the agent's identity, the memory's address level, and its expiration date, while keeping the actual memory content flexible. By standardizing these envelope fields and establishing a shared registry for semantic conventions, the industry could maintain tool-agnostic memory stores while allowing individual vendors to continue competing on the actual storage and processing logic.
While skepticism exists regarding the potential for another standard to simply add to the existing noise, the history of open-source progress suggests that fragmentation is overcome when the community merges existing efforts rather than launching isolated alternatives. Concerns about prompt injection and security are valid, but these can be managed through metadata that records the provenance of each memory. Addressing open questions about how to handle code dependencies, expiration defaults, and address hierarchies will require collaborative effort, but the path toward a unified memory spec is the necessary next step for maturing AI agent capabilities.
110 comments • Comments Link
- 现在就为 AI agents 制定标准化的内存格式还为时过早。这个领域仍处在实验阶段，而现有很多"内存"方案本质上只是把冗余内容塞进一套类似文件夹的结构，几乎无法在性能上超过模型的原生能力。
- 把 LLMs 当作黑箱来对待，会滋生"货物崇拜"式的工程：用户为了让 agent 的"内存"生效，发明出越来越复杂的框架；但实际上，模型本来就是为解析非结构化自然语言而设计的。
- 自动化的内存注入经常会引发"漂移"：agent 把那些短暂、只适用于特定任务的细节当成永久规则，进而导致幻觉增多、推理能力下降。
- 目前有一种很强的共识：手动整理通常优于自动化的内存系统。开发人员应当维护人类可读、便于审计的、面向"存储库级文档"的内容。
- 强行把 agents 塞进僵化的内存架构，往往会让性能变差。这说明模型在约束更少、结构更轻的时候反而能发挥得更好。
- 大量使用 LLMs 生成技术内容，已经让写作质量明显下滑：常见问题是内容含糊、重复、冗长，充斥"slop"，从而把真正的核心思想淹没掉。
- 在 agent 内存上推进企业级标准化，可能是一种"为了找问题而制造问题"的解法：它会带来不必要的复杂度、监控和官僚成本，却难以真正解决分布式系统状态这一核心难题。
- 可靠的 agent 工作流程通常依赖简单的基于文件的知识库：这类知识库能进行模糊搜索，也方便人工监督，而不需要复杂的数据库后端或专有的内存模式。
- 业界对 LLM 生成的文章普遍持高度怀疑态度。许多读者更偏好真实的人类声音，也对 AI 辅助写作里常见的表演感、以及"词汇沙拉"式的风格感到厌烦。
- 与其等待某个通用标准，不如让个人从业者通过构建自定义的轻量级检索系统来获得更直接的收益。他们能够针对具体使用场景亲自评估与调试这些系统。
这场讨论折射出一种矛盾：一方面，人们希望把传统软件工程的严谨性强加到 AI 工作流里；另一方面，LLMs 的优势恰恰在于在更灵活、且对人类更可读的语境中发挥作用。许多参与者都提醒不要过早标准化，并指出目前那些试图把"agent memory"形式化的做法，往往会引入无关或不准确的上下文，从而拖累模型表现。与此同时，大家对 LLM 生成的技术内容泛滥也感到明显不满——这些内容被认为质量本身就较低，且容易掩盖本该一眼就能抓住的简单概念。最终，社区更倾向于简洁做法：把项目特定的文档维护为标准文件。对于人类开发者及其 AI 同行来说，这仍然是管理知识最可靠、也最透明的方式。 • Proposing standardized memory formats for AI agents is premature, as the field is still in an experimental phase and current "memory" solutions are often just redundant folder structures that offer little performance gain over native capabilities.
• Treating LLMs as black boxes leads to "cargo cult" engineering where users invent complex frameworks for agent memory, even though models are fundamentally designed to parse unstructured natural language.
• Automated memory injection frequently causes "drift," where agents adopt transient, task-specific details as permanent rules, leading to hallucinations and the degradation of model reasoning.
• There is a strong consensus that manual curation is superior to automated memory systems, as developers should maintain repository-level documentation that is human-readable and easily audited.
• Attempting to force agents into rigid memory architectures often results in worse performance, suggesting that models are currently more capable when given minimal structural constraints.
• The widespread use of LLMs to generate technical content has led to a noticeable decline in writing quality, characterized by vague, repetitive, and overly verbose "slop" that obscures core ideas.
• The push for enterprise-grade standardization in agent memory may be a solution in search of a problem, potentially introducing unnecessary complexity, surveillance, or bureaucratic overhead without addressing the core challenges of distributed system state.
• Effective agent workflows often rely on simple, file-based knowledge bases that allow for fuzzy search and human oversight, rather than complex database backends or proprietary memory schemas.
• Industry skepticism toward LLM-generated articles is high, with many readers favoring authentic human voices and expressing fatigue over the performative, "word-salad" style common in AI-assisted writing.
• Rather than waiting for a universal standard, individual practitioners are finding more success by building custom, lightweight retrieval systems that they can personally evaluate and debug for their specific use cases.
The discussion reflects a deep tension between the desire to impose traditional software engineering rigor on AI workflows and the reality that LLMs function best with flexible, human-readable contexts. Many participants warn against premature standardization, arguing that current attempts to formalize "agent memory" often degrade model performance by introducing irrelevant or inaccurate context. There is a palpable frustration regarding the prevalence of LLM-generated technical content, which is viewed as inherently lower quality and prone to obfuscating simple concepts. Ultimately, the community consensus leans toward simplicity: maintaining project-specific documentation as standard files remains the most reliable and transparent way to manage knowledge for both human developers and their AI counterparts.