Learn · Context management & reliability · CCAR-F task 5.1

Context is a budget, not a bucket

Long sessions lose facts in the middle. Keep a sourced case-facts block outside the history, trim tool output, and load the rest just in time.

Key ideas

• Context is a finite attention budget: recall drops as it fills, and the middle is read least reliably.
• Keep critical facts in a short, structured, sourced case-facts block that lives outside the summarised history and is re-injected each turn.
• Progressive summarisation blurs exact amounts, dates and promises; keep a structured issue list for multi-issue sessions.
• Trim tool output to the fields you need; context editing can clear old tool results.
• Use notes files, just-in-time retrieval, /compact (with focus instructions) and /clear between unrelated tasks.
• For long inputs, put a key-findings summary first and use section headers.

Worked example: Compacting with a focus

In Claude Code, /compact keep the API design decisions summarises the session while telling the summary what must survive. Auto-compact does the same near the limit; in the API, server-side compaction plays this role.

Exam traps

• Answering "the agent forgets the order number" with a bigger context window.
• Asking the user to repeat facts the agent already verified.
• Relying on a running summary for exact amounts and commitments.
• Pasting every tool field and document up front instead of loading just in time.

Official docs

• Effective context engineering for AI agents
• Context editing
• Compaction

Practise CCAR-F questions · Next lesson: Escalation and ambiguity