CCA Prep
CCAR-FUA

Lesson 5.1Lesson 25 of 30

Preserving facts in long conversations

Manage conversation context to preserve critical information across long interactions

Preserving facts in long conversationsSummarised history only"Wants a partial refund soon"40+ fields per tool resultKey findings buried in the middleCase facts + trimmed contextCase facts block: $47.82 · 14 Mar ·#8812Tool output trimmed to 5 relevantfieldsKey findings first · section headersA bigger window does not stop summaries from blurring numbers

Key points

  1. 1Progressive summarisation blurs specifics: "refund of $47.82 by 14 March" becomes "customer wants a partial refund soon". Extract amounts, dates, order numbers, statuses and customer-stated expectations into a persistent case facts block included in every prompt, outside the summarised history.
  2. 2The Messages API keeps no state between calls: send the complete conversation history in every request. In multi-issue sessions keep each issue's structured data (order, amount, status, next step) in its own separate layer.
  3. 3Tool results accumulate and consume context out of proportion to their relevance: 40+ fields per lookup_order when 5 matter. Trim verbose tool outputs to the relevant fields before they pile up.
  4. 4Lost in the middle: the beginning and end of a long input are processed reliably, the middle may be missed. Put a key findings summary first and organise detailed results under explicit section headers.
  5. 5Subagents must return metadata (dates, source locations, method). When the downstream agent's context budget is small, upstream agents return structured facts, citations and relevance scores, not verbose reasoning.
What the exam guide checks here

Objective wording from the exam guide (English).

  • 5.1.K1Progressive summarisation blurs numbers, percentages, dates, customer expectations
  • 5.1.K2Lost in the middle: beginning and end are processed reliably, middle may be missed
  • 5.1.K3Tool results accumulate and consume tokens out of proportion to relevance
  • 5.1.K4Pass complete conversation history in each API request
  • 5.1.S1Keep a persistent case facts block (amounts, dates, order numbers, statuses) outside summarised history
  • 5.1.S2Persist structured issue data in a separate context layer for multi-issue sessions
  • 5.1.S3Trim verbose tool outputs to relevant fields
  • 5.1.S4Put key findings first and use explicit section headers
  • 5.1.S5Require subagents to include metadata (dates, sources, method) in outputs
  • 5.1.S6Upstream agents return structured facts, not verbose reasoning, when downstream context is limited

Lock it in

Answer scenario questions on this topic. Every option comes with an explanation.

Practise 5.1 · 26 questions