Gemini Memory Not Working: Check Settings and Scope

If Gemini memory is not working, check whether the problem is an unavailable feature, an unsaved detail, or a saved detail Gemini did not use. Then verify the account, activity settings, and conversation type against Google’s current help pages before changing anything.

For project work, keep important context in a record you can inspect. fenbs holds tasks, decisions, lessons, and progress outside a single AI chat. Gemini CLI and other supported assistants can connect to that record; the ordinary Gemini app does not connect to fenbs.

What is the quickest useful diagnosis?

1. Confirm the signed-in Google Account and whether it is personal, work, or school. Compare the account shown on each device.
2. Identify what should have carried over: an earlier chat, a remembered fact, an instruction, or a file you supplied.
3. Check Google’s current eligibility and supported-interface requirements for that exact feature.
4. Review the relevant Memory and Keep Activity controls, then test in an ordinary new text chat.
5. Save the observed result and change only one thing before repeating the test.

Use a harmless example, such as a preference for short summaries. Do not troubleshoot with a customer’s address, a private document, or a real deadline. The aim is to distinguish causes, so you need a test whose correct answer is obvious and whose failure creates no additional problem.

Are you expecting a feature this account supports?

Google’s past-chat memory requirements↗ (opens in a new tab) distinguish eligible personal accounts from work, school, and supervised accounts. They also require Keep Activity and identify supported interfaces and exceptions. Check that live page rather than assuming all Google accounts have the same controls.

Google separately documents remembered information and instructions↗ (opens in a new tab). That page has its own availability requirements. A control for written instructions does not establish that the account can save every fact through chat, and an unavailable fact-saving feature does not mean ordinary chat is broken.

If the documented requirements do not match your account, treat this as an availability boundary. Do not move organizational information into a personal account simply to obtain memory. Ask the administrator which approved workflow supplies ongoing work context, or provide a small permitted brief directly in each session.

Are memory and activity enabled for the right account?

Inspect the controls in the account that contains the original conversation. Check both the memory option and the underlying activity setting. A browser with several signed-in Google accounts can make two otherwise identical screens represent different histories, so confirm the profile before interpreting a missing chat.

Google’s Gemini activity guide↗ (opens in a new tab) explains where to review activity and manage Keep Activity. It also distinguishes personal-account controls from settings managed by a Workspace administrator. Review the implications of a setting before switching it on; memory convenience is only one reason someone may choose an activity policy.

Do not begin by deleting your history, resetting all preferences, or reinstalling the app. Those actions change the evidence and may remove information you still need. First establish whether the original detail exists in the intended account and whether the current session is allowed to use the relevant feature.

For team work, an explicit starting brief in fenbs can preserve the project’s goal and next action while you investigate. It also lets you record “activity setting checked; recall still failed” as a finding instead of asking the next assistant to reconstruct your tests.

Was the information saved, or merely discussed?

A fact appearing in a conversation is different from a fact being deliberately retained for future use. Ask the assistant what source it used, then compare its response with the original chat or visible saved information. “I will remember” is not a substitute for a test in a fresh session.

Keep the test narrow: “What response format did I ask you to use?” is clearer than “Tell me everything you remember.” If you paste the preference into the test prompt, a correct answer proves only that Gemini read the prompt. Test one known fact without supplying the answer again.

Could another context source explain the result?

A current attachment, a written instruction, an earlier chat, and a connected app can provide different versions of a fact. List the possible sources and ask which one informed the answer. For example, a draft brief may still contain an old delivery date after someone updated the agreed project plan.

Check the content before calling that a memory failure. Correct a bad source where you control it, and identify the approved source in your next request. Google warns that corrections and deletions do not always affect personalization immediately, so avoid declaring success from a single response directly after a change.

If the failure is confined to a specialized workspace or personalization feature, continue with Gems troubleshooting or personal context troubleshooting. The general AI memory hub explains other scope boundaries. Keep those investigations separate from this basic recall test.

How to do this with fenbs

1. Create a free account and add the project you need to keep moving.
2. Add the current tasks, a short context note, and the source of the approved requirements.
3. Write Where we left off with the failed recall test, checks already completed, and next step. Record an agreed rule separately in Decisions and rules.
4. Follow the integration guide for Gemini CLI or another supported assistant. Ask it to read the project context before beginning work.
5. Save progress and any verified lesson before you stop, including what remains uncertain.

fenbs does not repair Gemini memory or import your Gemini conversations. It is not the right fit if you only want personal preferences inside the Gemini app. Use it when you want a reviewed work record that supported assistants can read and update independently of that app’s recall.

What should you report if the test still fails?

Report the account type, interface, relevant settings, and exact steps using your harmless example. Say whether the control was absent, the fact was missing, or the saved fact was not used. Include the expected and actual result, without attaching private work as diagnostic evidence.

The same discipline helps with other tools: Copilot instructions troubleshooting separates supplied rules from recall, while ChatGPT connector troubleshooting separates access from content. An unrelated Claude coding failure needs its own diagnosis. Preserve the result that applies to this account and workflow.