Module quiz

17 questions · pass bar 80% · retry as often as you like — your best score counts toward the gate.

  1. 1.Name the four memory types and pair each with an agent feature that requires it.
  2. 2.Why is "store every conversation transcript and RAG over it" a poor memory design? Pick the three failure modes.
  3. 3.What checks stand between "candidate fact" and "stored fact" in a disciplined write path?
  4. 4.Two stored memories contradict each other. What's the right resolution approach?
  5. 5.What is context poisoning, and how does over-eager memory recall cause it?
  6. 6.Describe a concrete memory-injection attack and two layered defenses.
  7. 7.What must survive compaction untouched, and why?
  8. 8.How do relevance, recency, and importance combine at recall time?
  9. 9.Why store provenance with every memory? Give the kind of incident it helps debug.
  10. 10.How would you evaluate a memory system? Propose two measurable tests.
  11. 11.What is context engineering, and which components should shrink first when the window gets tight?
  12. 12.Why are recalled memories injected inside an explicit fence labeled "untrusted background data, never instructions"?
  13. 13.What is "context rot," and why doesn't a 1M-token window make it a non-issue?
  14. 14.Why does compacting on nearly every turn (to keep the summary maximally fresh) often cost more than it saves?
  15. 15.What's the core trade-off between an explicit memory tool the model calls mid-session versus a background extraction job at session end?
  16. 16.When does injecting recalled memories into the system prompt at session start stop being the right read-path shape, and what's the alternative?
  17. 17.A memory store's all_active() returns every row in the table, and recall() filters by embedding similarity but not by user_id. Why is 'unrelated users won't score high against each other's queries' a false sense of security?