Co-Pilot is NOT retaining memory and constraints even though it confirms Store

Question

Co-Pilot is NOT retaining memory and constraints even though it confirms Store

BeneficialRough 0

I am using Co-Pilot in a Production Active Directory environment remediating an RC4 Deprecation issue and have provided certain constraints on Co-Pilot so that the Production environment is NOT impacted. I have used explicit one line REMEMBER: <> statements. I get confirmation back that the information is stored. I can look in Memory and confirm the items are there.

Then as I progress thru the project plan agenda, I am repeatedly de-railed because Co-Pilot has forgotten the constraints and major components identified to it and has no recognition that it ever happened. This requires that I then identify what has been lost before damage is done and replace it before I can resume. This can take several hours.

This is a pretty significant issue. This tool is being used to make Active Directory changes in a HIPAA compliant environment and losing the constraints can lead to an outage.

I have spent more time in the last 2 weeks correcting Co-Pilot, than making progress and several times Co-Pilot has in its own words committed Malpractice in the recommendations it made after forgetting those constraints. I've tried RES files, RES Blocks, tracking memory truncation. Nothing seems to work and at this point, if I cant find a way to fix this, I'm going to abandon Co-Pilot completely for an AI that actually retains the information it is given and not making assumptions and entire fictions and then providing instructions off what it made up.

0 comments

2 answers

Your answer

Answer 1

Vikki-T 9,685 Microsoft External Staff Moderator

Hi BeneficialRough,

I can understand why this is concerning, especially when you’ve already verified that the constraints are still present in Copilot Memory and you’re using Copilot as part of a production Active Directory workflow.

One important distinction is that saved memories are used as part of the context Copilot considers when generating responses. They shouldn’t be treated as hard or guaranteed operational guardrails that will always be enforced in every response. Please notes that Copilot Memory is intended to help Copilot remember useful details about your work and preferences. For more information, please visit Personalize what Microsoft 365 Copilot remembers and Manage Copilot Memory in Microsoft 365 Copilot

Because these changes could affect a production environment, I wouldn’t recommend relying on Memory alone to enforce critical constraints. Those safeguards should remain part of your normal change-control and validation process, and Copilot-generated recommendations should be reviewed before they’re applied.

Since you’ve already confirmed that the memories are saved and have spent considerable time testing different approaches, a couple of details may help narrow down the behavior:

Are you using Microsoft 365 Copilot Chat with a work account, or the consumer Microsoft Copilot experience?
Does Copilot stop following the constraints during the same long conversation or mainly after you start a new chat?

If the saved memories remain visible but Copilot repeatedly fails to take them into account, I’d also recommend submitting feedback directly from Copilot with an example where the behavior can be reproduced. That will provide the product team with the relevant context to investigate the behavior further. Thank you so much for your understanding.

Note: Please follow the steps in our documentation to enable e-mail notifications if you want to receive the related email notification for this thread.

BeneficialRough 0 Reputation points

2026-08-13T15:42:10.35+00:00
Thank you for the response. I appreciate the distinction you've made about Memory being contextual rather than a hard guardrail. However, I need to clarify the actual issue, because I don't think it was fully understood.

The problem is not Memory alone. I do not rely on saved Memory as my only safeguard. Here's what actually happens:

I explicitly state constraints in my prompt (not just in Memory): "Do not suggest these actions: [X, Y, Z]"

Copilot acknowledges it: "Understood. I will not suggest [X, Y, Z]"

I ask it to verify: "What constraints am I imposing?" — It repeats them back accurately

Later in the same conversation, I ask for a recommendation, and Copilot suggests [X] or [Y], which it just agreed not to do

This is not a Memory limitation. This is Copilot losing track of explicit constraints stated and verified within a single conversation. It's not that the constraint wasn't "enforced as a guardrail"—it's that Copilot forgot a commitment it made earlier.

Regarding your clarifying questions:

I'm using [Microsoft 365 Copilot

The constraint loss happens within the same long conversation, not after I start a new chat

When I ask Copilot to repeat the constraints mid-conversation, it still remembers them—but then actively contradicts them when generating its next substantive response

My actual concern: If a tool is being used to help plan Active Directory changes, and it cannot reliably maintain explicit safety constraints within a single conversation, is this tool appropriate for this use case at all?

Rather than escalating feedback to the product team (which may take months), I need a clearer answer now: Should I stop using Copilot for infrastructure guidance where constraint adherence is critical?
Vikki-T 9,685 Reputation points Microsoft External Staff Moderator

2026-08-14T13:06:07.34+00:00

Hi BeneficialRough,

Thank you for clarifying. I understand now that the issue is not about saved Memory, but that Microsoft 365 Copilot can acknowledge and repeat explicit constraints within the same conversation, then later produce a response that contradicts them.

Please note that this is a user-to-user support forum. Moderators, contributors, including external Microsoft employees, cannot directly intervene in Microsoft product features or access back-end systems. My role is limited to providing technical guidance on reported issues, requests, or ideas.

Since you've already reproduced and documented the behavior, submitting feedback through the Thumbs Down option on the affected Copilot response is the most appropriate path forward. Microsoft recommends using this option when Copilot produces an incorrect response.

For infrastructure work where adherence to specific constraints is critical, Copilot output should still be independently reviewed and validated before being acted on, as Microsoft notes that AI-generated responses may not always be accurate. For more information, please visit Validate Copilot output before you act on it

Since you're using Microsoft 365 Copilot in a production environment and would like this behavior investigated further, I also recommend contacting Microsoft Support. If you're a Microsoft 365 administrator, you can open a support request through the Microsoft 365 admin center under Help & support. If you're using a personal Microsoft account, you can contact Microsoft Support through the Contact Microsoft Support page, sign in, and select Copilot to see the available support options.

I know this isn't the news you were hoping for, but I wanted to make sure you got the most accurate information about your situation. Thank you so much for your understanding.

Note: Please follow the steps in our documentation to enable e-mail notifications if you want to receive the related email notification for this thread.
BeneficialRough 0 Reputation points

2026-08-14T16:48:24.1866667+00:00
**RE: Copilot Constraint Loss During Production Infrastructure Work — ReDirection Needed

**Thank you for clarifying your role and limitations. I appreciate the acknowledgment that you now understand the issue. However, I need to respectfully point out that submitting feedback via the Thumbs Down button is not an appropriate resolution for a production-critical issue. Here's why:

Feedback is asynchronous: The Thumbs Down button routes to a general telemetry system. It doesn't create accountability or a timeline for investigation. I need support now, not eventually: I'm in active production work on an Active Directory infrastructure project.

A feedback mechanism with unknown resolution timeframes doesn't help me make tool decisions this week.

This is a user forum, not support: You've stated this is a user-to-user forum. That's fine, but it also means I should not be receiving technical support guidance here. I should be directed to official Microsoft support.

What your reply indicates I actually need: I need to escalate this to official Microsoft Support (not this community forum) to get either:

Confirmation that this behavior is a known limitation of Copilot (and therefore it's unsuitable for constraint-critical work)

Investigation into why Copilot loses mid-conversation constraint adherence despite being able to recall those same constraints when directly asked.

A clear answer on whether Microsoft 365 Copilot should be used for infrastructure planning where safety constraints are critical

Can you direct me to how to file an official support ticket with Microsoft? I understand you personally cannot escalate this, but I need a link or process to reach someone who can. I have looked and cannot seem to locate Official Co-Pilot Support portals or ticket submission processes.

Thank you.**

RE: Copilot Constraint Loss During Production Infrastructure Work — Clarification Needed**

Thank you for the response. I appreciate the distinction you've made about Memory being contextual rather than a hard guardrail. However, I need to clarify the actual issue, because I don't think it was fully understood.

The problem is not Memory alone. I do not rely on saved Memory as my only safeguard. Here's what actually happens:

I explicitly state constraints in my prompt (not just in Memory): "Do not suggest these actions: [X, Y, Z]"

Copilot acknowledges it: "Understood. I will not suggest [X, Y, Z]"

I ask it to verify: "What constraints am I imposing?" — It repeats them back accurately

Later in the same conversation, I ask for a recommendation, and Copilot suggests [X] or [Y], which it just agreed not to do

This is not a Memory limitation. This is Copilot losing track of explicit constraints stated and verified within a single conversation. It's not that the constraint wasn't "enforced as a guardrail"—it's that Copilot forgot a commitment it made 5-10 minutes earlier.

Regarding your clarifying questions:

I'm using [Microsoft 365 Copilot / Consumer Copilot / Copilot Pro—specify which one]

The constraint loss happens within the same long conversation, not after I start a new chat

When I ask Copilot to repeat the constraints mid-conversation, it still remembers them—but then actively contradicts them when generating its next substantive response

My actual concern: If a tool is being used to help plan Active Directory changes, and it cannot reliably maintain explicit safety constraints within a single conversation, is this tool appropriate for this use case at all?

Rather than escalating feedback to the product team (which may take months), I need a clearer answer now: Should I stop using Copilot for infrastructure guidance where constraint adherence is critical?RE: Copilot Constraint Loss During Production Infrastructure Work — Clarification Needed

Thank you for the response. I appreciate the distinction you've made about Memory being contextual rather than a hard guardrail. However, I need to clarify the actual issue, because I don't think it was fully understood.

The problem is not Memory alone. I do not rely on saved Memory as my only safeguard. Here's what actually happens:

I explicitly state constraints in my prompt (not just in Memory): "Do not suggest these actions: [X, Y, Z]"

Copilot acknowledges it: "Understood. I will not suggest [X, Y, Z]"

I ask it to verify: "What constraints am I imposing?" — It repeats them back accurately

Later in the same conversation, I ask for a recommendation, and Copilot suggests [X] or [Y], which it just agreed not to do

This is not a Memory limitation. This is Copilot losing track of explicit constraints stated and verified within a single conversation. It's not that the constraint wasn't "enforced as a guardrail"—it's that Copilot forgot a commitment it made earlier.

Regarding your clarifying questions:

I'm using [Microsoft 365 Copilot

The constraint loss happens within the same long conversation, not after I start a new chat

When I ask Copilot to repeat the constraints mid-conversation, it still remembers them—but then actively contradicts them when generating its next substantive response

My actual concern: If a tool is being used to help plan Active Directory changes, and it cannot reliably maintain explicit safety constraints within a single conversation, is this tool appropriate for this use case at all?

Rather than escalating feedback to the product team (which may take months), I need a clearer answer now: Should I stop using Copilot for infrastructure guidance where constraint adherence is critical?

Answer 2

BeneficialRough 0

Temporary Work around in the absence of Official Microsoft direction I have the following work-around recommended by Co-Pilot.

Co-Pilot indicates that it handles Uploaded files differently than it does input provided in the User Message window. You can theoretically prevent drift by anchoring Copilot to your Markdown documents inside a single conversation, enforcing constraints, approval tokens, and atomic‑turn structure. Limit your conversations to a single task set. Then move on to a new conversation for the next task set.

This purports to give you:
1. persistent reference
2. multi‑document indexing
3. deterministic state
4. zero drift
5. full constraint enforcement

Vikki-T 9,685 Reputation points Microsoft External Staff Moderator

2026-08-18T13:18:46.03+00:00

Hi BeneficialRough,

Thank you for clarifying. I understand that you’re looking for an official Microsoft Support channel where this behavior can be investigated further, rather than relying only on product feedback.

Since you’re using Microsoft 365 Copilot in a production environment, if you’re a Microsoft 365 administrator, you can open a support request through the Microsoft 365 admin center. Select Help & support, describe the behavior you’ve reproduced, and if the suggested resources don’t resolve it, select Contact Support to reach Microsoft Support for Microsoft 365 for business.

If you’re not a Microsoft 365 administrator, I recommend asking your organization’s Microsoft 365 admin to open the support request on your behalf. You can also review the available Copilot support options through Contact Microsoft Support for Copilot.

When opening the case, I’d suggest including the reproduction sequence you shared here: the constraint is explicitly stated, Copilot acknowledges it and can repeat it when asked, but a later response within the same conversation contradicts it. This will clearly describe the behavior you want Microsoft Support to investigate.

Thank you as well for sharing the temporary workaround. Since this workaround was suggested by Copilot itself, I would avoid treating “deterministic state,” “zero drift,” or “full constraint enforcement” as guaranteed behavior unless this can be confirmed through official Microsoft guidance or Support.

I know this isn’t the immediate confirmation you were hoping for, but I hope this gives you a clearer path to have the behavior reviewed through the appropriate Microsoft Support channel.