Why define expiry and pruning rules for agent memory?
Why B is the answer
Old facts can become wrong and mislead later work
Stale memory causes bad decisions, so memory needs expiration, pruning and reset rules.
In detail
- Facts go stale as a codebase changes.
- Old facts can mislead later work.
- Prefer facts that can be verified against the code.
Use case
An illustrative scenario, not a real incident.
An agent remembers an API path that was changed months ago, and it keeps using it. Expiry rules remove old facts before they mislead.
See it in different ways
flowchart LR
A["Long-running work"] --> B["Save progress to an artifact"]
B --> C{"Still matches the plan?"}
C -->|Yes| A
C -->|No| D["Correct, or reset from the artifact"]mindmap
root((Agent memory))
Short term
This task only
External
Files in the repository
Progress artifacts
Copilot Memory
Repository facts
Care
Expiry
PruningstateDiagram-v2 state "Fact stored" as s0 state "Fact used" as s1 state "Fact expired" as s2 [*] --> s0 s0 --> s1 s1 --> s0: still correct s1 --> s2: out of date s2 --> [*]
Key point: Old memories can mislead.
Why the other options are wrong
- A. Pruning makes each model respond in fewer than ten seconds
- Pruning is not about response time.
- C. Memory is billed per word on each and every request
- Billing per word is not the reason.
- D. Agents refuse to start if they hold more than ten memories
- There is no such limit of ten.