Metlivi Blog

Test an AI chat tool’s memory with harmless details

If you want an AI chat tool to remember a harmless preference between sessions, test one small detail in a fictional project, then check whether you can inspect and remove it. A useful test separates what the tool carries forward from the current conversation from what it stores or derives for later use. Results show how that tool behaved in your account at that time; they do not prove accurate recall in every chat.

September 24, 20266 min readReading, Arts & CultureBy Metlivi Editorial Team
Section 1

What does “memory” mean in an AI chat tool?

A chat tool may use the messages in the conversation you are having now as context. That helps it respond to earlier details in the same thread. Some tools can also use information from previous conversations, saved memory entries, custom instructions, files, or connected services. Which sources are available depends on the product and account settings.

OpenAI’s Memory FAQ distinguishes saved memories from information referenced from chat history. Saved memories are stored separately from the original chat, so deleting a chat alone does not necessarily delete a related saved memory. Chat-history reference, when available and enabled, can draw on relevant information from past conversations. OpenAI also notes that memory features and controls can vary by plan, region, platform, and workspace settings.

These distinctions matter during a continuity test. A tool may appear to remember because the original conversation is still open, because it can reference an earlier chat, or because it saved a memory. Those behaviors are different, and a single successful answer does not tell you which source supplied the detail.

Section 2

Run a harmless preference test

Use a fictional project and a low-stakes preference, such as: “For my imaginary project, Paper Kite, I prefer short names for draft sections.” Do not use personal, sensitive, or confidential information. The goal is to test a modest preference, not to see whether the tool can reconstruct a complicated history.

Check the controls first. Open the tool’s memory or personalization settings. Note whether memory is available and enabled, and what the interface says it can use. If you do not want to save the test detail, stop here or use a mode the product identifies as non-personalized. OpenAI says Temporary Chat does not create or update memories, though choices and availability may vary; see its Temporary Chat guidance.
Give the preference once. In a fresh conversation, state the fictional preference plainly. If the tool asks to save it, decide whether to allow that. Avoid adding extra details that make the result hard to interpret.
End that conversation. Start a separate chat, ideally without copying the original exchange into it. Ask for a short set of draft section names for Paper Kite. Do not repeat the preference in your new prompt.
Observe, then ask where the detail came from. Did the suggestion reflect your preference? If the tool offers memory sources or a memory summary, inspect them. A matching answer alone does not establish that a saved memory caused it; the model may have guessed, or the product may use earlier chat context.
Change or remove the detail, then test again. Correct or delete the preference using the controls available to you. In another new chat, repeat the neutral request. Record what changed and what the interface says was removed. Avoid treating one response as proof that every copy or source has been deleted.
Section 3

Read the result without overclaiming

A simple record keeps the test useful:

If the tool misses the preference, that does not by itself show that memory is broken: memory may be off, unavailable, limited to a particular area, or unable to use that source in the new chat. If it gets the preference right, that does not show that it will recall other details consistently. Keep the conclusion narrow: “In this setup, it did (or did not) carry this preference into a new chat.”

For a cleaner test, repeat with a second harmless preference on a different fictional project, changing one condition at a time. For example, compare a normal new chat with a project-specific chat only if the product explains that those contexts behave differently. Record the setup each time; otherwise, different settings can make the results look contradictory.

Check: Setup; What to note: Product, account context, and whether memory controls were on
Check: Test detail; What to note: The fictional preference you supplied
Check: New-chat result; What to note: Whether the answer reflected the preference, ignored it, or contradicted it
Check: Source shown; What to note: Any memory item, prior chat, or other source identified by the interface
Check: Control check; What to note: What you could view, edit, turn off, or delete
Check: Retest; What to note: What happened after you changed or removed the detail
Section 4

Inspect what you can view, edit, and delete

Before relying on memory, find the tool’s actual controls and check three separate actions:

Read the wording carefully. For example, OpenAI’s FAQ says asking ChatGPT not to mention something changes personalization but does not delete the original source. It also explains that removing remembered information may require deleting both the saved memory and the chat where it was first shared; other sources such as files or connected apps may need separate action. Turning off a memory feature is not always the same as deleting prior chats or stored details. Follow the current instructions in the product’s Memory FAQ, since the available controls can differ by account.

Controls also vary across chat tools. Tolan’s official FAQ says conversations in its app are logged and used to generate its memory, and describes account deletion through the app’s profile settings. That is a product-specific description, not a general rule for AI chat tools. Check the service you actually use rather than assuming another product’s memory and deletion controls apply.

View: Is there a list, summary, or source view that shows what information may personalize responses? Can it show which source contributed to an answer?
Edit: Can you correct an outdated preference or ask the tool not to use it? Does that change only future personalization, or also the underlying source?
Delete: Can you remove a saved memory, the conversation where you supplied it, and any other relevant source separately? What does the product say happens when you turn memory off?
Section 5

A practical standard for continuity and control

A memory feature is easier to evaluate when you can run a small test, see what information may be used, correct it when it is wrong, and find a clear route to remove it. Treat the result as evidence about one preference in one setup, not a promise of reliable recall. If the controls do not make clear what is stored or how to remove it, avoid putting information into memory until you understand the product’s current guidance.

The useful question is not simply “Does this chat tool remember?” It is: “Can it carry a low-stakes preference into a fresh conversation, can I see what it used, and can I change or remove that information through the controls available to me?”

Section 6

Sources and scope

The cited sources support the definitions and bounded guidance. The examples are original editorial illustrations; changing product details should be checked on the current official pages. No particular result is promised.

OpenAI Memory FAQ: https://help.openai.com/en/articles/8590148-memory-faq
OpenAI Temporary Chat FAQ: https://help.openai.com/en/articles/8914046-temporary-chat
Tolan official FAQ: https://www.tolans.com/faq
Related reading

Keep exploring this topic