Preserving critical facts in long conversations
How to keep amounts, dates, order IDs and rules intact as a conversation grows: pin facts in a case-facts block outside summarized history, trim tool output before it piles up, order long inputs for position effects, and choose between context editing, compaction and notes files.
Manage conversation context to preserve critical information across long interactions
~8 min read · 9 practice questions ·
What the exam expects
- Diagnose progressive summarization loss (exact amounts, percentages, dates and customer-stated promises turned into vague prose) and fix it with a persistent case-facts block and per-issue structured records kept outside the summarized history.
- Explain why a bigger context window does not remove the need to curate context (context rot), and why the stateless Messages API needs the full history on every request.
- Trim verbose tool results to the relevant fields before they accumulate, in the tool or in a PostToolUse hook.
- Assemble long multi-report inputs against the "lost in the middle" effect: key-findings summary first, each report in its own labelled section.
- Have upstream agents return compact structured data (key facts, citations, relevance, dates, sources) when the downstream agent's budget is tight.
- Choose between tool result clearing, compaction, notes files and CLAUDE.md for what must survive a long session.
Key terms: context rot (recall falls as tokens grow); case-facts block (transactional facts re-sent in every prompt, outside summarized history); progressive summarization (repeatedly condensing older turns); lost in the middle (findings in the middle of long inputs get omitted); tool result clearing (API context editing that replaces old tool results with a placeholder); compaction (replacing older turns with a Claude-written summary).
Key facts
- Detail
- Accuracy and recall degrade as token count grows, at any window size
- Detail
- Stateless; send the full conversational history every request
- Detail
- System prompt, all messages including tool results, tool definitions, output
max_tokens- Detail
- Caps generated output (up to 128k on 1M-window models), not input room
- Detail
trigger100,000 input tokens;keep3 tool uses;clear_tool_inputsfalse
- Detail
- Server-side; client keeps the full, unmodified history
- Detail
- Claude gets a warning near the threshold and can save results to memory files first
- Detail
- Precise recall of early details; exact state across many variables
- Detail
- PostToolUse
updatedToolOutput, any tool, both SDKs
- Detail
systemmessage with subtypecompact_boundary
- Detail
- CLAUDE.md via
settingSources, re-injected every request
- Why does a 1M-token window not remove the need to curate context?
- When should a fact go in the case-facts block rather than the conversation summary?
- Why pair tool result clearing with a memory or notes file?
- Where do persistent rules go so Agent SDK auto-compaction cannot drop them?
The questions are free to read. Answers with their sources and self-grading are included with the pack.
Keep reading with the Claude Certified Architect - Foundations pack
- Core concepts
- Exam traps
- Worked scenario
The pack includes:
- All 30 lessons of the Claude Certified Architect - Foundations guide, with reading progress
- Interactive decision walks and step-by-step sequence steppers
- Inline practice questions and one-click topic practice (9 questions for this lesson)
- An official source cited for every claim
- An AI tutor that answers from the guide
- Unlimited timed practice exams
Practice for the Claude Certified Architect - Foundations exam
Certifyforce is an independent study resource and is not affiliated with, endorsed by or sponsored by Anthropic. Claude is a trademark of Anthropic.