---
title: "Layers of memory — Context Field Notes"
description: "Working, session and long-term memory, and why they need different rules."
canonical: "https://press.getalchemystai.com/books/context-field-notes/memory-layers"
last_updated: "2018-10-20"
book: "Context Field Notes"
---

# Layers of memory

It is tempting to talk about "memory" as one feature. In practice there are at least three layers, and each one fails differently.

## Working memory

Working memory is the current request: the user's message, the tools that were just called, and whatever was retrieved a moment ago. It is small, precise and disposable. The main risk here is overflow. When a long tool output pushes the user's actual question out of view, quality drops sharply and silently.

> **Rule of thumb**
>
> Summarise tool output before it enters working memory, and keep the user's latest message last.

## Session memory

Session memory spans a conversation. It holds decisions that were made, preferences that were stated and facts that were corrected. Its main risk is contradiction: the user says one thing early, then another later, and both remain in view. A good session layer records corrections explicitly instead of hoping the model notices them.

## Long-term memory

Long-term memory outlives sessions. It is the most valuable layer and the most dangerous, because mistakes persist. Anything written here deserves provenance (where did this come from?), a timestamp, and a way to be forgotten on request.

The layers are not a hierarchy of importance. They are a hierarchy of commitment. The further down a fact goes, the more evidence it should need to get there.

---

- Previous: [Why context is the product](https://press.getalchemystai.com/books/context-field-notes/why-context.md)
- Contents: [Context Field Notes](https://press.getalchemystai.com/books/context-field-notes.md)
- Next: [Retrieval in practice](https://press.getalchemystai.com/books/context-field-notes/retrieval-in-practice.md)
