54 lines
1.7 KiB
Markdown
54 lines
1.7 KiB
Markdown
---
|
||
title: "State Externalization"
|
||
type: concept
|
||
tags:
|
||
- "harness-engineering"
|
||
- "agentic-ai"
|
||
- "memory"
|
||
sources:
|
||
- "Your-AI-Isn-t-Stupid---It-Just-Needs-a-Better-Harness--Lychee-Technology-Engineering-Blog"
|
||
last_updated: 2026-04-20
|
||
---
|
||
|
||
## Overview
|
||
State Externalization——将任务状态（pending/in-progress/completed）写入磁盘文件而非仅保存在 context window 的工程实践，是 [[Harness-Engineering]] 第二原则的具体实现。
|
||
|
||
## Core Principle
|
||
> 如果一条信息对任务连续性重要——已完成什么、待处理什么、什么失败了——它必须存在于 context window 之外。
|
||
>
|
||
> Context window 是易失的。磁盘文件是持久的。
|
||
|
||
## Two-Tier Memory Model
|
||
|
||
### Working Memory (Short-term)
|
||
当前步骤所需的即时对话和 context window。
|
||
|
||
### Persistent State (Long-term)
|
||
结构化文件（如 `state.json`）追踪精确的子任务状态：
|
||
```json
|
||
{
|
||
"task_id": "market-research-report",
|
||
"steps": {
|
||
"fetch_competitor_a": "COMPLETED",
|
||
"fetch_competitor_b": "COMPLETED",
|
||
"fetch_competitor_c": "FAILED",
|
||
"compare": "PENDING"
|
||
},
|
||
"last_updated": "2026-04-20T10:30:00Z"
|
||
}
|
||
```
|
||
|
||
## Why It Matters
|
||
- Context window 在每次新 LLM 调用时重置
|
||
- 状态不外部化 → 每次重置都丢失进度
|
||
- 进程崩溃 → 整个任务从头开始
|
||
|
||
## [[Context-Reset]] Integration
|
||
State Externalization 是 [[Context-Reset]] 的前提——必须先有可写磁盘的状态文件，才能实现有意义的 Context Reset。
|
||
|
||
## [[Memory-Consolidation]] Integration
|
||
State Externalization 写入的是原始状态，[[Memory-Consolidation]] 定期压缩这些状态防止膨胀。
|
||
|
||
## Source
|
||
- [[Your-AI-Isn-t-Stupid---It-Just-Needs-a-Better-Harness--Lychee-Technology-Engineering-Blog]]
|