Stop AI project context drift by giving every material file a version, effective date, owner and next review date; declaring which record wins when sources conflict; removing superseded files; and rerunning one representative output after any material change. Memory should help retrieve context, not decide what is authoritative.
Drift is an authority problem
Context drift appears when a project can still retrieve an old fact but cannot tell that it has been replaced. Typical symptoms include expired prices, competing calls to action, withdrawn claims, duplicated instructions, obsolete links and a connection that still exposes data the task no longer needs.
Use this precedence order:
1. Current approved source and claim register. 2. Current offer or subject canon. 3. Current project instructions. 4. Current asset brief. 5. Historical drafts and chat only as labelled evidence of prior work.
If two current records conflict, stop and ask for a decision. Do not let recency in chat silently settle authority.
Run a weekly exception scan
A lower-maintenance project does not need a full weekly audit. Scan only for exceptions:
- A source reached its review date. - The offer, price, promise, exclusion or canonical link changed. - A provider changed a feature, plan or permission surface. - A connection failed, widened scope or was no longer used. - A draft required material factual correction. - Outcome evidence crossed a pre-written continue, reshape, pause or stop threshold.
Run the full staleness review only when an exception appears or a risk-based due date arrives.
Test removal, not just addition
After updating a project, remove or archive the superseded file and run the old request again. The test passes only when the project no longer produces the retired price, claim, link or instruction. For a connection, revoke access and confirm that the project fails clearly rather than substituting invented data.
What repeated result should this project produce?
Keeps the project tied to one recurring job rather than a miscellaneous knowledge dump.
Which files are allowed to supply facts and which record wins when they conflict?
Establishes source precedence and exposes missing authority.
What happens if a plausible output is wrong or stale?
Sets the review and permission ceiling.
Can a named person inspect the result before it changes external state?
Separates a useful project from unsafe delegation.
Record the change once
Give the change an ID, affected canon version, affected assets, owner and completion evidence. Rebuild dependent outputs from the current authority. A one-off correction to a generated page hides the drift rather than repairing it.
Frequently asked questions
The project may still have access to an old file, conversation or connected source without a rule saying which record is authoritative. Remove superseded material and declare precedence explicitly.
Usually not. Run a brief exception scan weekly and trigger the full review only after a material change, failed output, permission event or risk-based due date.
No. Repeat the request that previously exposed the stale result. The repair is proven only when the retired fact no longer appears and missing authority produces an honest stop or question.
Clarify the right AI layer
The interview helps decide whether the recurring task still belongs in a project or should step down to a prompt, move into a controlled workflow or remain human-led.