Copilot Has a Bad Memory? Check Settings and Context

If Copilot seems to have a bad memory, first check which Copilot you use, which account is active, and whether the missing detail was saved or merely mentioned. Then test the relevant setting or source in a fresh conversation. Repeating “remember this” without those checks can hide the actual problem.

For work that must survive a chat change, fenbs keeps tasks, decisions, and progress in a deliberate record that connected assistants can read. That helps you distinguish a missing project fact from a preference the app did not apply.

What should you check first?

1. Name the app and account. Separate personal Microsoft Copilot, Microsoft 365 Copilot, and GitHub Copilot before following any troubleshooting steps.
2. Write down the exact missing fact and the expected answer. Use a harmless example rather than confidential work information.
3. Find where you supplied the fact: an old chat, saved memory, custom instruction, document, or coding repository.
4. Open the relevant setting or source and confirm the fact is present and current. Change one thing, then try a fresh conversation.
5. Record whether the assistant missed the fact, misunderstood it, or used an outdated version. Those are different failures with different fixes.

A useful test is “Which format do I prefer for summaries?” after deliberately saving that preference. A poor test is “Do you remember my project?” because almost any broad reply can sound plausible. Specific expectations make the result easier to judge.

Are you troubleshooting the right product?

Personal Copilot settings do not configure GitHub’s coding assistant. Likewise, a work account may have organization controls that your personal account does not. Keep the full product name and the signed-in account beside your test results so a browser profile change does not look like random forgetting.

The Copilot versus ChatGPT comparison helps identify the kind of assistant you are evaluating, while GitHub Copilot memory owns the detailed coding-memory explanation. This troubleshooting process focuses on locating the missing input.

Also distinguish a wrong fact from an absent fact. If an answer confidently repeats an old deadline, find and correct the old source. Adding the new deadline again without removing the contradiction gives the assistant two competing stories.

Is Microsoft Copilot using the intended preference?

For the current personal app, Microsoft places Saved memories under Settings, then Personalization. Review the relevant entry using its consumer privacy instructions↗ (opens in a new tab). Older screenshots may show a different route, so match the help page to your app.

In Microsoft 365 Copilot Chat, review saved entries through Chat settings, Personalization, and Manage saved memories. Microsoft’s memory-management page↗ (opens in a new tab) explains that control. Check the actual stored wording instead of using the assistant’s conversational promise as the only evidence.

Make the preference clear and bounded. “Use concise bullet points for my weekly summaries” is easier to test than “always write perfectly.” If the preference belongs only to one client deliverable, provide it with that task rather than treating it as a rule for every conversation.

For a stable response format in Copilot Chat, inspect custom instructions↗ (opens in a new tab) as well. A forgotten style request and a missing project approval should not be repaired in the same place.

Is GitHub Copilot missing instructions or current task context?

If the problem is in a coding workflow, check the repository you opened and the instructions that the particular feature supports. GitHub documents repository custom instructions↗ (opens in a new tab) separately from learned memory. A consumer personalization toggle will not repair a missing repository file.

Compare the instruction with the behavior you expected. Does it state the relevant build command, file convention, or approval boundary clearly? Remove obsolete alternatives and confirm the current source is the one you intend. Our Copilot instruction examples help with wording without turning today’s task progress into a permanent rule.

If you use the GitHub Copilot app, check its documented configuration for that experience too. Do not assume that an instruction you wrote for another surface automatically reached this session. Ask the assistant what inputs support its plan, then inspect those inputs yourself.

Was the detail ever a dependable project record?

A chat can contain abandoned ideas, corrections, and unfinished work in the same thread. Asking an assistant to “remember the discussion” leaves it to decide which statement matters. Write the final fact explicitly: what was approved, who approved it, which task it applies to, and what remains unverified.

For example, “The smaller form was approved; message delivery still needs testing” preserves a distinction that “the form is ready” loses. When the next chat misses the delivery check, you can point to a clear record instead of reconstructing the conversation.

The fenbs context workflow gives connected assistants the Rules, current work, and Where we left off at the beginning of a session. Ask the assistant to read that record before treating its recollection as the plan.

How to do this with fenbs?

1. Create a free account and add the project whose context keeps getting lost. One person can start free, with no credit card.
2. Add the current task, record the decisions a person made, and write the specific fact or check that a new session must not miss.
3. Follow Integrations to connect a supported assistant over MCP, a standard way to use an outside tool. fenbs supports GitHub Copilot; it does not connect to Microsoft Copilot or Microsoft Teams.
4. Ask the assistant to read the project context and state its next step before working. At the end, ask it to save progress and any remaining uncertainty.

fenbs will not fix a broken app setting, and it does not import your old Copilot conversations. If a single personal style preference is all you need, manage it inside the app. Use fenbs when the missing information is the work itself.

When should you report a recall problem?

Report a reproducible problem after you have confirmed the right app, account, setting, and source. Include a harmless test prompt, the expected result, the actual result, and whether it happens in a new conversation. Leave confidential project information out of the report.

One failed answer does not identify a permanent defect, and one successful answer does not prove every future handoff is safe. Keep checking important facts against their source. You can find more explanations of recall and context in the AI memory hub.