Agent Memory as a File Format

Memoryfield是一种将AI Agent内存表示为可移植文件格式的新方法1。该格式采用ZIP文件结构，包含Markdown页面和可选的SQLite向量索引1。单个内存页面设有约8KB的软限制，相当于约2000个token1。

这一设计通过语义搜索而非图遍历来检索信息，最多需要2次工具调用，相比传统方法的N+1次调用更加高效1。该方案推荐采用nomic-embed-text-v1.5作为嵌入模型1。Memoryfield支持多种传输方式，包括本地文件、Amazon S3、GitHub和HTTP等1，并提供了命令行工具和Agent技能供用户快速开始使用1。作者还发布了示例Memoryfield（soapstones.memoryfield.zip）用于演示1。

A new approach to AI agent memory management has been introduced through Memoryfield, a file format that represents agent memory as portable, transferable data rather than processing logic.1 The format packages memory as a ZIP file containing Markdown pages paired with an optional SQLite vector index, enabling efficient information retrieval through semantic search rather than graph traversal.1

The design addresses limitations in existing agent memory systems by treating memory as data that can be accessed across different platforms and tools.1 Individual memory pages operate under a soft limit of approximately 8KB, equivalent to roughly 2,000 tokens, and semantic search reduces information retrieval to a maximum of two tool calls compared to the N+1 calls required by traditional graph-based approaches.1 The implementation recommends nomic-embed-text-v1.5 as the embedding model and supports multiple storage and transport methods including local files, Amazon S3, GitHub, and HTTP.1

The format includes command-line tooling and agent skills to facilitate adoption, with a demonstration example provided as soapstones.memoryfield.zip.1