A freshness gate is proposed to tag each memory chunk with its age and a per-class TTL, enforcing retrieval rules before data reaches the model: STALE_BLOCK for expired, STALE_WARN for near-expiry with down-ranked relevance, and FRESH for still-current facts.

Query scenarios show naive top-k can confidently output stale data, whereas freshness gating yields deterministic behavior by refusing or downgrading stale facts, preventing misleading answers.

In a concrete example with near-duplicate price chunks, the older price tag (e.g., $29) can be blocked and the newer, fresher chunk ($39) used when it is still current, while the older one loses top-1 status.

Readers are encouraged to share the shortest-lived fact ever quoted with confidence and to calibrate TTLs for their domain, emphasizing that TTLs should be configurable and domain-dependent.

A real-world AI memory issue arises when near-duplicate, age-invariant embeddings cause stale but correct facts to be recalled with high confidence due to identical similarity scores and tiny margins.

A code appendix provides a deterministic script (freshness_gate.py) that models the gate with fixed reference time and TTLs, demonstrating naive versus gated retrieval and showing STALE_BLOCK, STALE_WARN, or FRESH outcomes along with the injected token or refusal.

Mechanism details include stored_on timestamp, volatility class (price, availability, reference, etc.), per-class TTLs, and a similarity floor that filters out off-topic chunks, with a freshness score mapping age to a 0-1 scale.

Practical recommendations include stamping stored chunks with stored_at and source lineage, categorizing by volatility class, inserting the gate between retrieval and the model, using it for audit trails to explain usage or refusal, and adding gates for live fetches to avoid garbage content.