An agent must recall user preferences across separate sessions. Which memory type fits?
Why D is the answer
Long-term memory in an external store
Context only lives for the session, so durable facts need external storage.
In detail
- Preferences across sessions need long-term memory in an external store.
- Let users see and delete what is stored.
Use case
An illustrative scenario, not a real incident.
A travel assistant remembers a user's seat preference next month, because it is saved outside the session.
See it in different ways
flowchart TD
A{"How long must it last?"} -->|"This task"| B["Short-term: state and context"]
A -->|"Across sessions"| C["Long-term store"]
C --> D["Episodic: what happened"]
C --> E["Semantic: facts"]
C --> F["Procedural: how to act"]mindmap
root((Memory))
Short term
Current context
Summaries
Long term
Episodic
Semantic
Planning
Plan and execute
Revise on new factsmindmap
root(("Agent memory"))
n1["Session"]
n2["Context window"]
n3["Across sessions"]
n4["External store"]Key point: Durable facts need external storage.
Why the other options are wrong
- A. The tokenizer's session cache
- A tokenizer holds no memory.
- B. The context window of each session
- A context window ends with the session.
- C. The KV cache on the serving GPU
- A KV cache is for generation.