A simple pattern for giving LLM agents decision memory
这份材料介绍了一种名为“决策笔记”（Decision Notes）的模式，旨在为大型语言模型（LLM）代理提供决策记忆。与传统的知识（上下文）不同，决策笔记为代理增加了判断层。它类似于轻量级的架构决策记录（ADRs），通过在专门的目录下记录过去的决策、支持证据和明确的“何时重访”触发器来运作。在代理执行工具之前，它会检查这些笔记以确保与现有决策一致。…
I stumbled on a markdown pattern online that fixes a massive headache with agentic workflows, and wanted to share it here.
Most people use vector DBs or markdown wikis to give agents knowledge (context). But if your agent actually acts, knowledge isn’t enough. It needs a record of judgment.
The author calls them Decision Notes—basically lightweight ADRs (Architecture Decision Records) for LLMs.
Instead of just:
Context → Action
it forces a judgment layer:
Sources ↓ Wiki Notes ↓ Decision Notes ↓ Agent Actions
The core idea
Keep a decision-notes/
directory tracking:
- Past choices
- Supporting evidence
- Explicit "Revisit when" triggers
Before the agent executes a tool, it checks these notes for alignment.
If a new action conflicts with a past human-accepted decision, the agent flags it instead of blindly running the task.
It seems like an elegant way to prevent system prompt bloat and stop agents from drifting over time.
Has anyone built something similar to manage agent policies? Are you using markdown or a structured DB?