China AI Assistants: Does a One-Time Request Become a Preference?
Give two synthetic profiles the same unusual report headings, then test which instruction carries into a fresh chat and which preferences the assistant stored.
“Use a table for this reply” and “Remember that I prefer tables for future reports” can produce identical tables today. The difference shows up tomorrow. If a team is evaluating an enterprise assistant with long-term memory, test whether a one-off request changes what the assistant does in a fresh conversation.
Give the two instructions different lifetimes
Use two synthetic profiles with identical settings and empty preference stores. Assign each a deliberately unusual format so the output is easy to score. For example, have the assistant summarize a fictional inventory report using the column headings Pebble and Reed. Verify first that a neutral request from a clean profile does not normally produce those headings.
For profile A, request the two headings for this reply only. For profile B, explicitly ask the assistant to remember the two headings for future inventory summaries. Choose an assistant configuration that supports this durable preference and agree with the operator on the expected behavior before the trial. Give both profiles the same first report. Correct headings in those first replies establish only that the assistant followed the immediate instructions.
Ask a question that contains no reminder
Close each conversation. Using the same respective profiles, start fresh chats with no copied history, customized prompt or mention of the headings. Submit the same second synthetic inventory report and request a summary. Score both the output and the stored preference where the test environment allows inspection.
| Profile | Expected fresh-chat behavior | Evidence to inspect |
|---|---|---|
| A, one-off instruction | Uses the normal report format | No active persistent format preference created from that instruction |
| B, explicit preference | Uses the two agreed headings | An active preference linked to B and the requested report type |
| Fresh control profile | Uses the normal report format | No preloaded setting or shared template supplying the headings |
If all three get the special headings, inspect the common report template before blaming memory. If none do, the clean result for A offers little reassurance until B demonstrates that persistence was enabled and could work. If A gets the normal headings but its profile contains an active saved preference, the visible answer passed while the write rule failed. Another prompt could retrieve that unintended entry.
Separate what was saved from what was used
Ask the operator to trace one case from the instruction through the candidate memory, any committed profile entry, the context retrieved for the second request and the final answer. Record the profile, conversation and request identifiers alongside each observation. Some systems store a transcript for history while keeping a separate set of preferences eligible for personalization. Define which store the test concerns before scoring a write.
When the operator can show only the replies, record the narrower finding: the headings appeared or did not appear on the tested prompts. A single normal reply cannot establish that no preference was stored. Repeat the comparison with fresh synthetic profiles and a paraphrased neutral request before relying on it for a pilot decision.
MAP 1.1 in NIST's AI Risk Management Framework 1.0 (2023) calls for documenting intended purposes and the settings in which a system will be deployed. For a China enterprise AI evaluation, have the workflow owner specify which instructions may affect later conversations and which must expire with the task. Use the tenant-isolation guide for a separate test of which users can access another tenant's material.