DSP(数据结构协议):如何为大型存储库中的 LLM 代理提供“长期记忆”

📄 中文摘要

在与代理工作时,常常会发现一个模式:最初的 5 到 15 分钟并不用于任务本身,而是用于“熟悉环境”。在大型项目中,这种熟悉过程不仅令人烦恼,还会不断消耗代币和注意力。DSP(数据结构协议)通过将项目地图外部化为一个简单、版本化、语言无关的图形,使其与代码并存,并作为持久内存供代理使用。该协议旨在提高代理的效率,减少在项目中寻找入口点和依赖关系所需的时间和精力。

📄 English Summary

DSP (Data Structure Protocol): how to give LLM agents "long-term memory" of a large repository

A common pattern recognized by those working with agents is that the first 5 to 15 minutes are often spent not on the task itself, but on getting oriented. This orientation process can be particularly burdensome in large projects, leading to a constant drain on tokens and attention. The Data Structure Protocol (DSP) externalizes the project map into a simple, versioned, language-agnostic graph that exists alongside the code, providing persistent memory for the agent. The goal of this protocol is to enhance the efficiency of agents by minimizing the time and effort required to locate entry points and dependencies within a project.

Powered by Cloudflare Workers + Payload CMS + Claude 3.5

数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等