Why does AI keep using an old decision?

An earlier chat, memory entry, project file or retrieved source may still contain it. Ask for the supporting source and compare it with your current approved record before assuming which mechanism caused the answer.

Does a source citation prove a decision is current?

No. A citation helps you locate evidence; the evidence may describe an earlier decision. Check its approval or effective date and whether another record replaces it.

Does marking a Multiplist record CURRENT update every assistant automatically?

No. CURRENT is a label you add to the saved record. Connect and authorize the assistant, ask it to retrieve that record, and verify the returned content. Other memories, instructions and files may also need correction.

Will deleting an old chat remove its remembered information?

Not necessarily. ChatGPT saved memories can exist separately from chats, and Claude's current memory experience also keeps generated memory entries after a source chat is deleted. Review the relevant memory controls separately.

What if the assistant cannot retrieve the current record?

Confirm that the record was saved and that the correct account and connector are enabled. If retrieval still fails, provide the reviewed text directly and check the answer against it.

Why Is AI Using an Old Decision? How to Update Its Context

Direct answer: An assistant may be drawing on an earlier chat, memory entry or reference file. Find the supporting source, save an explicit current decision, and check that the next answer uses it. Multiplist gives your connected tools a shared record of the knowledge you choose to carry forward.

You changed the plan. Your assistant is still confidently drafting from the old one.

You should not have to argue with an earlier version of your own work. The useful question is: which decision did it find, and what shows that decision is still in force?

Multiplist, your portable memory vault, gives you a place to keep the current decision alongside its supporting material. A connected assistant can retrieve that saved source when you continue the work. You still choose what to save, identify what is approved, and check what was retrieved.

Add Multiplist to Claude

The link opens connector setup with Multiplist's details filled in. Complete sign-in and authorization, enable the connector where you work, and save the material you want to reuse. Connecting the vault does not collect all your previous chats. See the connection guide.

# First, ask what the answer is based on

Before asking for another rewrite, try:

Which source supports the decision you just used? Show its title, date or version, and the passage you relied on. If you cannot find a source, say so.

Open the returned source and read the passage. An assistant's statement that it remembers your decision is not enough to establish where it came from. If it cannot locate evidence, treat the cause as unresolved and supply the current record.

A citation can be accurate and still point to an outdated decision. It answers where this came from; you must also check whether this is the decision to use now.

For example, an October 2 planning conversation might say that a workshop will be recorded. An October 10 approval changes it to a live discussion with a written follow-up. Both records can be useful history. Only one describes the current delivery plan.

# Save the current decision, with the change made explicit

Save a reviewed note or document in Multiplist. Give it a specific title so you can ask for it again. A short record is enough:

Title: Workshop delivery — CURRENT — approved October 10, 2026 Decision: Run a live discussion and send a written follow-up. Do not promise a recording. Replaces: The recording plan in “Workshop planning — October 2, 2026.” Include that source's link or ID. Why: The approved format prioritizes participant discussion. Next step: Update the invitation using this delivery plan.

This is an illustrative record, not a claim about a real workshop. Use the actual approval, effective date and supporting source for your own work. A newly uploaded copy of an old document is not automatically a newer decision.

Keep the earlier source when its history matters. Where your tools allow it, add a pointer explaining which record replaces it. Multiplist supports saving sources and retrieving a specific saved source; its tool reference describes those operations.

CURRENT and Replaces are labels you put in the record. They do not make every assistant select that record automatically, and they do not override its other instructions. Their purpose is to make your intended decision clear enough to verify.

# Retrieve the record before continuing the task

In your next connected conversation, ask:

Retrieve “Workshop delivery — CURRENT — approved October 10, 2026” from Multiplist. Show the saved decision and the earlier plan it replaces. Then draft the invitation using the approved delivery plan. If you find conflicting sources, identify them before drafting.

Check the returned title and content against the saved note. If several records have similar names, use the exact source ID or link to distinguish them. Confirm the actual decision, rather than relying on a search result's short description.

Then check the output: does the invitation promise a written follow-up, without offering a recording? That small test tells you whether this answer followed the record you supplied. It does not prove that every future conversation will retrieve it.

If retrieval fails, confirm the source was saved, the correct account is connected, and the connector is enabled for this conversation or task. You can also provide the reviewed text directly while resolving access. Important work should not depend on an unverified recollection.

# Correct native memory and project references where needed

Native memory can be enough when the correction belongs within one assistant and you can verify it there. You do not need a separate vault for every preference. Multiplist is useful when you want to maintain a reusable decision and its sources across connected tools.

In Claude: In the current memory experience, open Settings → Memory → Topics to review and edit a topic, or tell Claude what to change in chat. Projects have separate memory spaces. Paid chat search respects project boundaries and can cite original conversations. Organization controls affect availability. See Claude's memory guidance.

In ChatGPT: Open Settings → Personalization → Memory. Correct outdated information in chat or use the available summary or saved-memory controls. Options vary by account, region, platform and workspace. A summary, or Sources beneath an answer, may not show every contributing detail. See OpenAI's memory guidance.

Deleting a chat is not a reliable shortcut: ChatGPT saved memories can remain separate, and Claude's current memory experience retains generated entries after their source chat is deleted. Review memory separately. Claude's legacy experience differs; use the guidance for the controls you actually see.

Also inspect the project instructions and reference files you supplied. Correcting a memory entry does not revise a separate document containing the old plan.

Product facts checked October 10, 2026.

# Carry the correction forward

Once the record is right, use its name or source ID at the start of the next task. Ask for the decision, its supporting evidence and any unresolved conflict before generating more work.

For a different supported assistant, check its current connection options through Multiplist. Authorize the connection and confirm it can retrieve your chosen source. If it cannot, share the reviewed record manually. Saving content involves Multiplist's hosted service; review the privacy policy when choosing what to store.

Tags: ai-memory · cited-memory · claude · chatgpt · shared-source-of-truth · All Learn