Govina Labs 的开发者在 Hacker News 上发布了名为 CheeseDB 的开源嵌入式内存数据库,该工具专为 AI Agent 的记忆管理设计 1。CheeseDB 采用 Go 标准库编写,无外部依赖,其核心特性是“来源感知”能力,即每条记忆记录都关联其信息来源 1。当用户删除某个不可信或错误的信息源时,所有基于该源构建的记忆可被自动级联清除 1。这种机制允许开发者选择 Fair 模式(保留有独立备份来源的记忆)或 Strict 模式(删除所有关联记忆),从而彻底擦除潜在污染数据 1。
除了逻辑上的删除功能,CheeseDB 还支持物理擦除操作,能够直接从磁盘文件中移除文本字节,并通过哈希链技术检测数据篡改 1。为了保障多 Agent 环境下的安全性,系统通过 Token 实现不同 Agent 间的数据隔离 1。在 Apple M1 Max 硬件平台上进行的基准测试显示,处理 10,000 条记忆的 Recall 操作时,CheeseDB 的性能比 Neo4j 快约 13 倍,比 SQLite 快约 45 倍,而相比 PostgreSQL 则快约 370 倍 1。
目前,CheeseDB 处于 Preview 阶段,尚未接受独立安全审查,且官方声明 API 和磁盘格式可能在 v1.0 版本前发生变更,因此不建议将其用于生产环境中的关键数据 1。用户可以通过 Go 库调用、CLI 命令行接口使用该软件,也可借助 Model Context Protocol (MCP) 将其集成到 Claude Code、Cursor 等 AI 编程助手中 1。
Govina Labs has released CheeseDB, an open-source embedded memory database engine designed specifically for AI agents. 1 Written in Go using only the standard library with no external dependencies, the tool functions as an append-only, hash-chained memory store that prioritizes source-aware data management. 1 Its core innovation lies in linking every memory record to its specific sources, allowing users to delete unreliable or corrupted information—such as poisoned web content or erroneous instructions—and automatically cascade-delete any memories derived from those sources. 1
The system offers two distinct deletion modes to handle complex data relationships: "Fair" mode preserves memories that have independent backup sources, while "Strict" mode removes all associated records regardless of other connections. 1 To ensure data integrity and security, CheeseDB supports physical erasure that removes text bytes directly from disk files, detects tampering through its hash chain structure, and isolates different agents using token-based authentication. 1 In performance benchmarks conducted on an Apple M1 Max chip, processing Recall operations for 10,000 memories proved approximately 13 times faster than Neo4j, 45 times faster than SQLite, and 370 times faster than PostgreSQL. 1
CheeseDB integrates with various development environments, supporting usage via Go libraries, CLI commands, and the Model Context Protocol (MCP) for connections to AI coding assistants like Claude Code and Cursor. 1 Despite its capabilities, the project remains in a Preview stage; developers note that it has not undergone independent security audits, and both the API and disk format may change before the v1.0 release. 1 Consequently, the team advises against using CheeseDB for critical production data at this time. 1
评论
还没有评论,欢迎留下第一条。