What Is a Context Window?

What does context window mean?

A context window is the maximum amount of text or other tokenized input available to a large language model at one time when generating output, usually measured in tokens rather than words or characters, per Wikipedia's definition of the term. Everything sent to the model in a single exchange, the system prompt, the conversation history, any retrieved documents, and the response the model generates, all share this one limited space. Anything outside that window is invisible to the model unless it gets summarized, retrieved again, or reintroduced explicitly.

The practical effect is straightforward: a model cannot act on information it cannot currently see, no matter how well it handled that same information three messages ago.

Why the context window matters for product managers

Builders Camp's AI Agents bootcamp lists memory and context management as one of its core skills, framing it as designing "context windows, retrieval, and state so agents stay consistent." That framing matters because an AI agent working through a multi-step task, researching, drafting, revising, accumulates information with every step. Tool outputs, intermediate reasoning, and retrieved documents all consume the same limited context space as the original instructions.

A PM who does not account for this will watch an agent behave reliably for the first few steps of a workflow, then start ignoring earlier decisions once the context window fills and older content gets pushed out or summarized away.

How the context window is used in practice

Builders Camp's Building with Claude Code bootcamp names this failure directly as "memory rot": a certification quiz answer describes it as what happens "when long sessions accumulate outdated decisions, abandoned approaches, and contradictory instructions" inside the context window, with the recommended fix being to start a fresh session with a short rehydration briefing and clear source documents, rather than trying to cram an ever-growing history into the same window indefinitely.

In practice, this means a PM building an agent-based workflow has to decide deliberately what belongs in the context window at every step, and what should instead live in an external, retrievable memory that only gets pulled in when relevant. That decision is the actual skill, not the size of the window a given model happens to offer.

How Builders Camp teaches the context window

The AI Agents bootcamp covers context window management as part of its planning and orchestration module, alongside the human-in-the-loop and evaluation practices needed to keep a multi-step agent reliable over a long task. Building with Claude Code extends the same idea into its own context architecture module, treating a structured, reusable context system as the fix for context windows that would otherwise fill with one-off instructions repeated every session.

For a hands-on example of managing context in a real integration, see Build an AI Assistant with MCP, or read about vector stores as a way to retrieve only what a context window actually needs. See the AI Agents bootcamp for the full curriculum.