Data freshness in an AI agent is how old the information behind a decision is allowed to be before the agent must refresh it, flag it, or stop. It is not the same as "latest available." A stock level that is ten minutes old can be fine for a weekly report and wrong for placing an order.
Atlan describes the closely related idea of context freshness as "the practice of keeping the definitions, lineage, ownership, and policy an AI agent relies on synchronized with the current state of the systems those things describe" (Atlan, July 2026). A freshness policy turns that practice into rules an agent can check before it acts.
Why stale data is an agent problem
A person reading a stale dashboard can notice that a number looks wrong. An agent acts on it. It quotes an expired price, ships to an old address, or approves a payment against a balance that changed a minute ago.
Retrieval makes this harder to see. A document embedded last quarter keeps answering questions long after the policy it describes was replaced, and nothing in the answer says so unless you record the source date.
Freshness contract
| Field | Example question |
|---|---|
| Source of truth | Which system can confirm the current entitlement or balance? |
| Maximum age | Minutes for a dashboard, seconds for an execution decision? |
| Effective time | Does the record apply now or only after a future date? |
| Invalidation | Which event expires cached or embedded data? |
| Failure action | Use labeled stale data, request review, or stop? |
Set the window by decision, not by system
The same record needs different freshness depending on what the agent does with it. Write the window next to the decision, not on the data source.
| Decision | Freshness it usually needs | If the data is too old |
|---|---|---|
| Answering a policy question from documents | Days or weeks, with the document date shown | Answer, and show the source and its date |
| Quoting a price or availability | Minutes, re-read before sending | Re-read the source, or hand to a person |
| Changing a record, sending money or shipping goods | Re-read immediately before acting | Stop and record why |
RAG and execution differ
Retrieved guidance may tolerate a declared age. A high-impact tool action should re-read authoritative state immediately before execution when balances, permissions, inventory, or destinations can change.
For retrieval, re-index when the source changes rather than on a fixed schedule, and store the effective date with every chunk so an answer can carry it. For execution, never let a cached value stand in for a re-read on a consequential step.
What to do when data is stale
- Label and continue, when the decision is low impact and the age is shown to the user.
- Refresh and retry, when the source of truth is reachable within the time budget.
- Route to a person, when the source is unreachable and the decision matters.
- Stop and record the reason, when acting on stale data could move money, change records or reach a customer.
Evidence
Record source, record ID, version or checksum, retrieval time, effective time, cache state, freshness decision, policy version, and outcome. Test stale, future-dated, cross-tenant, duplicated, and contradictory records.
Preserve freshness evidence without copying sensitive source content. For where those records live and how long to keep them, see AI agent audit trail retention.