AIMentor — AI Expertise Learning Platform

13 questions · pass mark 80% · 20 minute limit. The timer auto-submits at zero.

Time remaining: 20:00

Question 1easy
Why is memory a design problem rather than a free feature?
LLMs have unlimited contextThe context window is finite and billed per token, so you must decide what the model sees each turnModels remember everything permanently after trainingMemory only matters for images
Question 2easy
What is the tradeoff of simple window/trim short-term memory?
It is expensive and slowIt is cheap and predictable but forgets anything that scrolls out of the windowIt never forgets anythingIt requires a vector database
Question 3medium
How does summarization memory preserve old context at bounded cost?
By deleting the system promptBy compressing older turns into a running summary while keeping recent turns verbatimBy embedding the whole conversationBy increasing the context window size
Question 4medium
Long-term memory is best implemented as:
Keeping the entire conversation in the message list foreverStoring facts externally and retrieving only the relevant ones per turn (memory as RAG)Fine-tuning the model after each messageRaising the temperature
Question 5hard
Why does retrieval-based long-term memory scale to thousands of facts without bloating context?
Because it compresses all facts into one tokenBecause only the memories relevant to the current turn are retrieved and injected, not all of themBecause the model memorizes themBecause vectors take no space
Question 6medium
On a big multi-step task, why can stuffing every step's output into one agent's context hurt?
It makes the model fasterIrrelevant detail piles up, cost climbs, and focus drifts ('context poisoning')It improves faithfulnessIt has no effect
Question 7hard
What is context isolation via sub-agents?
Running the same agent twiceDelegating a subtask to a separate agent with its own fresh context that returns only a distilled resultDisabling tools for the main agentStoring context in a database
Question 8hard
After a sub-agent completes, what should the orchestrator receive?
The sub-agent's full message history including every tool callOnly the distilled result/summary; the intermediate work is discardedNothing — sub-agents can't return valuesThe sub-agent's system prompt
Question 9medium
Match strategy to fit: which are correct? (Select all that apply)
Windowing — simple chats where old context is safely disposableSummarization — long single conversations where the gist must persist at bounded costRetrieval — facts that outlive the conversation or exceed any windowSub-agent isolation — a single one-shot prompt with no steps
Select all that apply.
Question 10medium
Which is a valid privacy concern for long-term memory?
It makes answers longerIt stores user data, so you must support deletion and scope it per userIt cannot store textIt disables retrieval
Question 11medium
What does a typical production assistant combine?
Only a fixed 2-message contextA summarized short-term window + retrieved long-term memories + sub-agents for heavy subtasksOnly fine-tuningA single infinite context
Question 12medium
Long-term agent memory is essentially the same technique as which Track-3 concept (acronym)?
Question 13hard
What technique (two words) delegates a subtask to a separate agent with its own context to avoid polluting the main one?

0 / 13 answered

Memory & Multi-step Design — Final Exam