Step 1 of 82 · Host app → Modeltask + 5 toolsTurn 0 of 12 · context 0 of 2,000 tokens · 0 tokens in all · history: 0 papers

Chapter 5 of 7

Memory, and what survives it

In chapter 4 the host stopped a stuck loop, and every turn re-sent the whole history, so each one cost more. Now Monday’s storm brings five damaged orders at once. The history grows with every one, and one customer has asked for something the agent must not forget.

This chapter: What does an agent remember from turn to turn, and what is lost when the host cuts its history down?

You will learn

Why the context grows with every result the agent gets back
What compaction saves, and what a summary can drop
How pinning keeps the line that has to survive

Words you’ll meet (4)

Context window: How much text a model can read in one request, counted in tokens (a token is about ¾ of a word). You met it in RAG, chapter 2.
Context budget: The most tokens this host lets one turn send the model. Larkspur’s team set it; it is not the size of any model’s context window.
Compaction: The host replacing older history with a short summary, so later turns send less.
Pinning: Keeping a line outside the summary: the host sends it every turn and never compacts it.

The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.

Step 1 of 6

A new case: Monday’s storm soaked a delivery van. Five customers say their orders arrived damaged: 2101, 2102, 2103, 2104 and 2105. For each one, ship a replacement if we have stock, otherwise refund it, and email the customer. One order, 2101, carries a note from Fenwick Farm Supply: “Email only, please. No phone calls: our phone line is down until November.” The host keeps the whole history and re-sends it every turn, as in chapter 4. It also sets a context budget: one turn may send the model at most 2,000 tokens. That number is Larkspur’s setting for this host, not the context window of any model. Before you run it:

Five orders, and every result kept. Will each turn fit in the host’s budget of 2,000 tokens?

The case

The task

Monday’s storm soaked a delivery van. Five customers say their orders arrived damaged: 2101, 2102, 2103, 2104 and 2105. For each one, ship a replacement if we have stock, otherwise refund it, and email the customer.

The agent’s tools

get_order, check_stock, create_replacement, issue_refund, send_email

The host’s rules

A context budget of 2,000 tokens per turn. Larkspur’s team chose it for this host; a model’s own context window is a different, fixed limit. Memory: full keeps every result; summary compacts the history once it passes 900 tokens; pinned compacts too, and keeps customers’ notes in a pinned line.

Counting tokens

Bytes of text ÷ 4, the rule from RAG chapter 2. Source: OpenAI Help Center, What are tokens and how to count them?

Context is a finite resource with diminishing returns: as it fills, the model's ability to recall information from it accurately decreases. Source: www.anthropic.com (checked 2026-10-20)