Context Compaction: How AI Agents Survive Long Sessions

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-foundations-context-compaction-r7a4-creators

1. What inherently happens to information each time an AI agent performs context compaction?

   1. Compaction removes only redundant information and leaves no gaps in memory
   2. The agent gains additional storage capacity by converting data to a new format
   3. All previous conversation history remains perfectly intact and accessible
   4. Some information is inevitably lost during the compaction process

2. In the context of agent memory, what does the term 'eviction' specifically refer to?

   1. Deleting all user messages from the conversation history
   2. Relocating stored memories to a different physical server
   3. Removing tool-call noise while keeping the outcomes and decisions that resulted from those calls
   4. Forcing the agent to forget its original system instructions

3. Before an agent performs compaction on a long conversation, what three categories of information should it explicitly classify as must-survive?

   1. Previous compaction dates, memory locations, and processing times
   2. All tool definitions, API keys, and system prompts
   3. User preferences, conversation topics, and emotional tone
   4. Current goal, key constraints, and recent error context

4. What does 'file-based offload' enable an AI agent to do?

   1. Compile code into a standalone executable program
   2. Store artifacts in external files and retrieve them by reference rather than keeping them in active memory
   3. Delete all files on the user's computer to free up space
   4. Transfer conversation history to a different AI agent for processing

5. Why is it recommended that agents make compaction visible to users rather than performing it silently?

   1. To prevent users from asking too many follow-up questions
   2. Because visible compaction improves the agent's processing speed
   3. To comply with legal requirements around AI transparency
   4. So users can identify if important information was lost and provide corrections

6. What capability can true long-context models provide that compaction strategies cannot substitute for?

   1. Built-in fact-checking for all historical statements
   2. The ability to run without any internet connection
   3. Faster response times than standard context models
   4. Genuine access to all information without any summarization loss

7. What aspect of a conversation does recursive summarization primarily aim to preserve?

   1. All emotional expressions and their intensity levels
   2. The exact timestamps of all interactions
   3. Every single word spoken by every participant
   4. The narrative flow and key story elements across the conversation

8. What happens when an AI agent 'forgets' an earlier instruction that a user provided?

   1. The agent switches to a different personality to compensate
   2. The conversation is automatically terminated for security reasons
   3. The agent automatically asks the user to repeat the instruction
   4. Users typically feel betrayed because they expect the agent to remember everything

9. An agent is having a 3-hour conversation with a user. What problem will it eventually face without compaction strategies?

   1. It will begin generating responses in a different language
   2. It will automatically terminate the conversation to protect itself
   3. Its responses will become permanently slower regardless of hardware
   4. It will exceed its context window limit and be unable to process new inputs

10. What type of details cannot be recovered by an AI agent, even with advanced compaction techniques?

    1. Data that exceeds one megabyte in size
    2. Conversations that occurred in a different session
    3. Details that the agent never explicitly recorded in the first place
    4. Information that was deleted more than 24 hours ago

11. Which of the following is NOT a compaction strategy mentioned in the content?

    1. Recursive summarization
    2. File-based offload
    3. Episodic eviction
    4. Recursive encryption of conversation history

12. What is the primary purpose of building a compaction prompt around the must-survive classification?

    1. To make the user feel more confident about the agent's abilities
    2. To ensure critical information is explicitly preserved during the compression process
    3. To allow the agent to ignore user instructions
    4. To speed up the overall compaction operation

13. In the eviction strategy, what specifically gets removed while outcomes are preserved?

    1. All punctuation and formatting from messages
    2. The user's original query text
    3. Tool-call noise, meaning the technical details of how tools were executed
    4. User authentication credentials

14. If an AI agent uses compaction perfectly but still produces incorrect results, what fundamental limitation is at play?

    1. Compaction was never actually performed correctly
    2. The agent is deliberately trying to produce wrong answers
    3. The user did not provide enough initial information
    4. Compaction inherently loses some information, which may include details needed for accuracy

15. What distinguishes file-based offload from simple deletion of information?

    1. Offload preserves the information and allows it to be retrieved by reference later
    2. Offload automatically optimizes the file size for faster processing
    3. Offload permanently destroys the information after a set time period
    4. Offload converts text to audio format for accessibility