An agent that depends on an external vector store for long-term memory experiences a scenario where the vector store is unavailable due to a planned maintenance window. The agent is mid-task when the outage begins. Which degraded-mode behavior MOST appropriately handles this situation?
Select an answer to reveal the explanation.
Short Explanation and Infographic
An agent that can only function when all its dependencies are up is fragile in production. Graceful degradation means: keep running with what you have (in-context memory), queue writes for later (don't lose new information), and be transparent about the limitation (mark output as degraded). Stopping cold wastes all in-progress work. Asking the user to substitute for your memory system is not graceful — it's offloading your infrastructure problem onto the human.
Full explanation below image
Full Explanation
Graceful degradation in agent systems requires defining a reduced-capability mode that preserves forward progress without compromising output integrity, and being transparent about what is and is not available.
Option A (terminate and wait) is the simplest response but the most wasteful. If the agent is 70% through a task and the memory store goes down for a 10-minute maintenance window, terminating and restarting wastes the completed work and delays delivery. In most tasks, in-context memory is sufficient to complete a session — long-term memory retrieval happens primarily at the start of a task, not continuously throughout it.
Option B is correct because it implements four-part graceful degradation: (1) Continue with in-context memory — for the current task, what the agent has already retrieved into context remains available. New retrievals that would require the vector store are skipped (or noted as 'retrieval unavailable'). (2) Buffer new writes — memory entries the agent would normally write during the task are queued in a local durable buffer (e.g., a file, a database). This ensures no new information is lost. (3) Flush on recovery — when the vector store comes back, the buffer is flushed, and all queued writes complete. (4) Degraded-mode notice — the task output is tagged to indicate it was produced without full long-term memory access, so human reviewers know to verify any decisions that might have benefited from historical context.
Option C (ask user for context) shifts the infrastructure problem onto the user. It is appropriate only if the missing context is critical and cannot be reasonably inferred from what is available in-context. Making this the default degraded-mode behavior makes the agent dependent on human knowledge substitution for every memory store outage.
Option D (cache entire vector store locally) is impractical for large stores (potentially gigabytes of embeddings) and creates a consistency problem — the local cache diverges from the live store during the agent's session. Large vector stores cannot practically be cached in memory on an agent instance.