Question 40
Domain 5A team attaches 10 internal docs (~20 KB each) to a Claude.ai Project. After the first week, users complain Claude 'forgets' earlier turns much sooner than expected. What is the most likely structural cause?
Correct answer: D
Explanation
Claude.ai Projects load attached documents into the working context for each chat, so those 10 internal docs consume part of the model’s context window. With less room left, the conversation history and new answer have a smaller shared budget, so earlier turns drop out sooner and Claude appears to “forget” them.
Why each option is right or wrong
A. The Project is silently summarizing earlier turns to make room.
Summarization may happen in some systems, but the core issue here is fixed context consumed by attached docs.
B. The model is operating below its context limit but above its 'attention budget,' a separate cap.
A separate "attention budget" cap is not the standard explanation; context window usage is the relevant constraint.
C. Claude.ai's tier-based cache eviction is dropping older turns under load.
Tier-based cache eviction describes infrastructure behavior, not the normal cause of shorter conversational memory in one chat.
D. The 10 docs are loaded into context every chat, shrinking the budget the conversation and answer share.
Claude.ai Projects place attached files into the active prompt for each conversation, so the 10 internal documents are not just stored passively; they occupy part of the model’s fixed context window on every chat. Under the context-window limit, once those ~200 KB of material are included, there is less remaining space for prior turns and the next response, causing older messages to be truncated earlier than users expect.