Your AI Agents Keep Forgetting Everything. Google Just Changed How They Remember. 你的 AI 智能体总是忘记一切？Google 刚刚改变了它们的记忆方式

TL;DR

• Google Cloud introduced the Open Knowledge Format (OKF) v0.1, an open spec for representing business knowledge as plain markdown files with YAML frontmatter, eliminating the need for custom context pipelines per project
• OKF is built on three simple conformance rules: parseable YAML frontmatter, a non-empty type field, and reserved filenames following expected structure—deliberately low-barrier to encourage adoption
• The format uses one-file-per-concept, markdown links as an untyped knowledge graph, and is designed to be git-native, tool-agnostic, and portable across agent frameworks
• Third-party tooling is already emerging, including okc (Postgres-to-OKF) and claude-okf (a slash command that auto-generates bundles from codebases using Claude)
• OKF v0.2 is already in progress, adding provenance, trust, lifecycle, and attestation as first-class concerns while maintaining backward compatibility

Why It Matters

The "Karpathy problem"—where wikis rot because humans neglect them—has long been a bottleneck for AI agent adoption, forcing every team to build fragile, framework-locked context pipelines. OKF offers a portable, human-readable standard that collapses custom ingestion layers into a single write-once, use-everywhere bundle, potentially becoming the shared knowledge layer that agent ecosystems have lacked. For practitioners, this means agent context can finally be decoupled from tooling choices and maintained alongside source code.

Technical Details

• Format structure: An OKF bundle is a directory of markdown files with YAML frontmatter; each file represents one concept (e.g., BigQuery table, API endpoint, metric, runbook). The only required frontmatter field is type, used for routing and filtering.
• Knowledge graph via markdown links: Relationships between concepts are expressed as standard markdown links (e.g., [deploy-rollback.md](/runbooks/deploy-rollback.md)). Graph-view consumers treat all links as directed, untyped edges; the relationship semantics live in the surrounding prose.
• Conformance rules: Three hard rules—parseable YAML frontmatter, non-empty type field, reserved filenames (index.md, log.md) follow expected structure. Everything else is soft guidance; consumers must tolerate unknown types, missing optional fields, broken links, and extra keys.
• Index and log files: index.md provides progressive disclosure by listing directory contents; log.md records dated, newest-first change history per directory so agents can detect recent changes without full diffing.
• Reference implementations: The knowledge-catalog repository includes a reference agent for auto-generating bundles, a graph viewer, and worked samples (GA4 Google Merchandise Store, Stack Overflow schema, Bitcoin dataset). Third-party tools like okc (Postgres schema → OKF bundle) and claude-okf (slash command for Claude Code) are already built on the spec.
• Versioning strategy: OKF v0.1 is a starting point; v0.2 adds provenance, trust, lifecycle, and attestation. Minor bumps are backward-compatible; major bumps may rename required fields or change reserved filenames.

Industry Insight

• End of per-project context pipelines: Teams should migrate from building custom parser-transform-agent chains to adopting OKF-style bundles, reducing engineering overhead and framework lock-in when switching agent runtimes.
• Knowledge as code: OKF treats business context as versioned, diffable, git-native artifacts—enabling agents to coexist with CI/CD workflows, code reviews, and collaborative editing rather than living in isolated, decaying wikis.
• Emerging ecosystem opportunity: With v0.2 adding trust and attestation, OKF is positioning itself as a foundational layer for enterprise agent deployments where provenance and data lineage matter; early adopters who build OKF tooling or migrate their knowledge bases now will have a first-mover advantage as the spec matures.

TL;DR

• Google Cloud推出Open Knowledge Format (OKF) v0.2，一个将企业知识编码为纯markdown文件（带YAML frontmatter）的开放规范，解决AI agent上下文管理的碎片化问题
• OKF采用"一个概念一个文件"设计，通过YAML frontmatter提供结构化元数据（type字段为核心），用markdown链接构建知识图谱，实现跨agent框架的互操作性
• 兼容性要求极低：仅需三条规则（可解析的YAML frontmatter、非空type字段、保留文件名遵循结构），允许未知类型和断裂链接，确保bundle在重构过程中保持可用
• 配套提供参考agent（自动生成OKF bundle）和图查看器，第三方工具如okc（Postgres schema转bundle）和claude-okf（Claude Code slash命令）已开始基于规范构建
• 核心价值主张是"Write once, use everywhere"：知识源只需生成一次bundle，任何agent框架均可直接读取，消除重复开发自定义上下文管道的成本

为什么值得看

这篇文章揭示了AI agent落地企业场景的核心瓶颈——模型能力已足够，但缺乏共享的企业知识格式，导致每个团队重复构建定制化上下文管道。OKF提供了一个低门槛、高兼容性的开放规范，让企业能够将分散在wiki、Slack、代码库中的业务知识转化为agent可消费的标准化格式，显著降低agent开发和维护成本。

技术解析

核心架构与文件格式：OKF bundle是一个markdown文件目录，每个概念（BigQuery表、API端点、指标定义、incident runbook等）对应一个独立文件。YAML frontmatter包含必需字段type（用于路由和过滤）及推荐字段title、description、resource（规范URI）、tags、timestamp。markdown链接作为知识图谱的有向无类型边，关系语义由周围文本承载。

兼容性规范：conformance bar刻意保持低位——bundle只需满足三条规则即可兼容：非保留markdown文件含可解析YAML frontmatter、frontmatter含非空type字段、保留文件名（index.md、log.md）遵循预期结构。消费者必须容忍缺失可选字段、未知类型值、未知额外键、断裂跨链接和缺失index文件，确保bundle在重构、增长和agent编辑过程中保持可用。

配套工具与生态：repository包含参考agent（自动扫描代码库生成OKF bundle）和图查看器。第三方工具如okc将Postgres schema转换为带外键作为图边的cross-linked bundle；claude-okf在Claude Code中实现/okf slash命令，自动扫描仓库概念、生成markdown文件、运行合规审计，并通过hook检测bundle与代码的漂移。

版本演进方向：v0.1为起点，已演进至v0.2，新增provenance（溯源）、trust（信任）、lifecycle（生命周期）和attestation（验证）作为一等公民关注点，同时保持核心格式的非强制性。小版本添加向后兼容字段，大版本可重命名必需字段或更改保留文件名。

行业启示

Agent基础设施标准化趋势加速：OKF的出现标志着AI agent领域从"每个项目定制上下文管道"向"共享知识格式"演进，类似数据库领域SQL标准化的路径。企业应停止为每个agent框架重复开发上下文管道，转而采用开放规范降低长期维护成本。

企业知识资产化新范式：OKF将分散在wiki、Slack、代码库、senior engineer头脑中的隐性知识转化为机器可读、agent可消费的标准化格式，使企业知识成为可版本控制、可跨团队共享、可自动维护的资产，而非依赖个人记忆的脆弱载体。

低门槛采用策略：三条兼容性规则意味着团队可从单个概念文件起步，逐步扩展bundle规模，无需一次性重构全部知识体系。建议AI从业者立即在小规模场景试点OKF，验证agent上下文管理效率提升，同时关注v0.2新增的信任与溯源机制对企业级部署的影响。