Agent memory

When you will meet it

When the agent has forgotten everything from yesterday's conversation, this is why: the context window is the desktop, memory is the drawer, and they are not the same thing. Miss the distinction and you will think a bigger context window cures forgetting — a bigger desk does not by itself move things out of the drawer.

Minimal example

沒有記憶：
  每次 session 從零開始；「我上次叫你別用那個 API」不存在

有記憶（示意）：
  寫入側：任務結束 → 把決策、教訓、用戶偏好寫進檔案或資料庫
  讀取側：新任務開始 → 用關鍵詞或向量檢索挑出相關幾條 → 塞進上下文

本站的三層設計：
  Session Log（原始對話記錄）→ Daily Log（當日整理）→ MEMORY.md（長期精華）

Both sides are hard, differently: the write side struggles with when to write and at what granularity — left to the agent's goodwill it skips logging when busy, and this site documented an incident where over ninety percent of content never reached the structured log; the read side struggles with which entries to pick — wrong picks are worse than none, injecting stale or irrelevant content as current truth.

What people get wrong

• Conflating memory with the context window. Context is everything the model sees this session and dies with it; memory is durable storage outside, appearing only when actively fetched and injected. RAG puts documents into context by the same re-injection mechanism.
• Solving only the write side, or only the read side. Store everything and inject all of it every time and you have no memory plus a bigger bill; the quality bottleneck of a memory system is usually selection.