An agent that manages issue triage loads its context from a memory store at startup. The memory store contains information about repository state, open issues, and team assignments that was cached 72 hours ago. The agent makes triage decisions based on this stale data, incorrectly assigning closed issues and referencing team members who have since left the project. What is the root cause?

Memory freshness is a critical property in agentic systems that operate over time. Context that was valid when it was cached can become incorrect as the environment changes — issues get closed, team members change, priorities shift. Without a freshness policy, the agent loads and trusts old data as if it were current, leading to decisions based on facts that no longer hold.

A freshness policy defines a maximum age (TTL — time to live) for each category of cached context. When context is loaded, its timestamp is compared to the current time. If the context is older than the TTL, the agent should either: (1) re-fetch the data from the live source before proceeding, or (2) annotate the loaded context as potentially stale and reduce its decision weight, or (3) abort and require a context refresh before continuing.

For highly dynamic data like issue state and team assignments, the TTL may need to be measured in minutes or hours, not days.

Option A (model hallucination) does not fit the scenario. The agent is making decisions consistent with the stored data — the data itself is outdated, not fabricated.

Option C (expired token) would prevent live API calls but would not explain why the memory store contains incorrect data about people who 'have since left.' Token expiry causes API errors, not stale memory content.

Option D (storage capacity exceeded) would cause context to be truncated or missing, not to contain wrong values about specific issues or people.

Freshness timestamps and TTL policies are mandatory for context stores in production agentic systems.