Agent Memory Systems Quiz

Answers and explanations appear as you take the quiz.

1. Why add external memory to a stateless LLM request?
   • Information outside the current context must be stored and retrieved if a later request needs it
   • Every model has identical fixed memory limits
   • A larger prompt automatically preserves information forever
   • External memory guarantees perfect recall
2. Which memory type stores facts and learned preferences about the user or world?
   • Procedural memory
   • Semantic memory
   • Episodic memory
   • Cache memory
3. What metadata does each Generative Agents memory object include?
   • Only a vector with no associated record
   • A fixed summary with no time information
   • A natural-language description, creation timestamp, and most-recent-access timestamp
   • A user ID and a guaranteed truth label
4. According to the topic, what is the genuinely hard part of memory systems?
   • Buying enough disk space
   • Choosing a font for the output
   • Translating memories into binary
   • Curation: deciding what to remember, consolidate, retrieve, and forget
5. What does a 'reflection' step typically do in a memory system?
   • It periodically distills raw logs into higher-level facts or summaries
   • It deletes the entire database nightly
   • It doubles the context window size
   • It encrypts every memory
6. Which user control is useful when an agent stores personal memories?
   • A promise that stored facts are always correct
   • A way to inspect, correct, or delete stored facts under an explicit policy
   • Automatic sharing of all memories with every agent
   • An assumption that users never need to manage memory
7. Which factors does the Generative Agents retrieval function combine when ranking memories?
   • File size, token count, and alphabetical order
   • Only semantic similarity
   • Relevance, recency, and importance
   • Only the most recent record, regardless of query
8. What should a memory system do when stored facts conflict or become stale?
   • Always prefer the newest statement without review
   • Keep every conflict hidden from the user
   • Treat retrieved memory as verified truth
   • Track source and time, then update, qualify, or remove information under an explicit policy