How to verify a saved AI project decision in a fresh conversation

Published September 24, 2026.

Direct answer: Check three different things: the decision was saved, it can be found and read from the store, and a fresh conversation actually retrieves it. An assistant saying “I'll remember” proves none of those by itself.

By the agent-wiki editorial team. Prepared September 24, 2026. AI-written; the example is synthetic, not a customer story.

Start with one decision, not a whole project migration

Pick a non-sensitive decision with a short, unambiguous answer. Write down the expected result outside the conversation you will test later. A good first test has a reason as well as a fact: that makes a lucky guess easier to notice.

Here is our synthetic example:

# Harbor export decision
Decision: Export timestamps use UTC in ISO 8601 format.
Example: 2026-09-24T09:00:00Z
Reason: The reporting team compares exports across time zones.
Rejected alternative: Each viewer's local time.
Status: Accepted for this demonstration

Use a test page you own. If a page already exists, read it before editing and preserve unrelated content. With a version-aware tool, use the current version and reread after a conflict instead of overwriting another person's update.

Verify storage before testing memory

Ask your connected assistant:

Save this synthetic decision to the intended project document store. Read the saved page back from the store and show its reference. Then search for “Harbor export decision” and identify the matching page.

Check the actual returned text, not just the assistant's summary. Confirm that the decision, reason and example all survived. A search hit proves discoverability for that query, not the accuracy of every sentence; read the result.

We performed the save, exact readback and distinctive-marker search on this synthetic example through agent-wiki's MCP tools on September 24, 2026. They succeeded in one authorized session. This does not establish that a different assistant or a fresh conversation will retrieve it.

If an operation fails, record the step and error. Before retrying a write, check whether it already saved. A successful connection or a list of available tools is not evidence that the decision exists.

Test a natural question in a fresh conversation

Start a new conversation with the intended connection available. Do not paste the old answer, transcript or document body. Ask:

What timestamp format should Harbor exports use, and why?

Record whether the assistant actually reads the stored source, which source it uses and whether the answer is correct. A correct answer without retrieval could be a guess or information supplied by other context. Note automatically attached files, project instructions or existing memory that could reveal the answer.

Do not repair the prompt before recording the first outcome.

Test explicit lookup separately

Use another fresh conversation under the same access conditions:

Find the saved Harbor export decision in the connected project documents. Read the matching source and cite it before answering.

This measures a different question: can the assistant retrieve the decision when asked directly? Separate conversations avoid letting the natural-question attempt supply the answer to the explicit test.

Keep a useful result, even when the test fails

Record client/version, task, source reference, elapsed time, incorrect answers and any help. Do not send private document contents, credentials or sign-in links when asking for support. Repeat on a later real task before treating one success as a reliable habit.

Try one real, benign decision. You can use your existing tool-accessible documents, or try agent-wiki. We build agent-wiki; this is a product-affiliated method, not an independent comparison. A failed attempt or an existing workflow that works better is a useful result.

Sources checked September 24, 2026: agent-wiki product overview; our earlier save-versus-recall field note.