- Nov 13
Copilot Memory and Agents: Review the Boundary Before Work IQ Expands Access
- IAMPathway Trainer
- Microsoft Entra ID
An agent can respect your Microsoft 365 permissions and still learn more about you than its business task requires. Copilot memory makes that distinction important: information used to personalize your assistant is not automatically information every approved application needs.
A public archive of Message Center notice MC1539004 reports a planned Work IQ MCP capability for agents to read the signed-in user's memory profile. It describes a December 2026 rollout and a default-enabled memory-read policy. Those specific details remain reported plans, not independently verified released behavior in this article.
What Microsoft documents publicly is already enough to justify a review: Copilot memory exists, Enhanced personalization is on by default, and Work IQ combines delegated authorization with tenant policy. Here is how to evaluate that boundary without confusing a permissive setting with universal agent access.
What is verified—and what is still a reported plan
As of November 7, 2026, Microsoft's memory administration guidance describes saved memories, custom instructions and details inferred from chat history. It marks personalization and memory as preview and subject to change. It also confirms that Enhanced personalization is enabled by default.
Separately, Microsoft's Work IQ MCP governance documentation confirms layered authorization: Microsoft Entra authentication, OAuth permissions, the user's existing permissions and request-time tenant policy. Policy cannot create access the user does not have.
The public third-party archive of MC1539004 reports a more specific expansion:
A proposed MCP resource path,
/memory/me/coreMemories.Read access through the
fetchtool, with no memory writes through that path.Access to the authenticated user's profile, not other users' profiles.
A memory-read policy enabled by default.
Worldwide general availability planned to start in early December 2026 and finish in late December.
Respect for personalization settings, with changes potentially taking up to 48 hours.
These claims come from an unofficial mirror, not a Message Center notice authenticated in a tenant during this review. The official roadmap URL for item 578869 returned the general roadmap page without the item's details in the public content retrieved. Searches of public Microsoft material did not independently establish the memory path or its rollout schedule.
That is an evidence gap, not proof the announcement is wrong. Before acting on those particulars, check the current notice in your own Message Center and its linked documentation. Do not treat December as a contractual deadline, or the proposed path as a working production endpoint today.
Two defaults must stay separate: Microsoft Learn confirms default-on Enhanced personalization. The memory-specific MCP read-policy default is reported by the archived notice. One does not independently prove the other.
Memory is more than a preferred writing style
Microsoft describes Copilot memory as including saved memories, chat-derived details and custom instructions. Its Enhanced personalization explanation describes personalization using communications, work data, usage signals, profile information and connector data.
That does not mean a memory response contains every email, transcript or file. A source used to derive personalization is different from the contents of a returned profile. The final memory-path schema and exact fields were not verified here.
For security teams, the useful distinction is between three kinds of context:
Explicit instructions: what a user has asked Copilot to do, such as preferring brief responses.
Saved details: information retained for reuse rather than supplied again in each conversation.
Inferred context: details learned from previous interactions and, where applicable, work activity.
The archived announcement describes custom instructions, saved memories and chat-history-derived memories at launch. Treat broader claims about exactly which work-account fields an agent receives as unconfirmed until the published schema establishes them.
An illustrative example: a user might ask an assistant to remember a preference for plain-language summaries. Another might discuss a confidential project repeatedly. Both are personalization context, but they have different sensitivity. These are hypothetical examples, not observed responses from the proposed API.
The governance question is therefore not simply, “Can this app read work data?” It is, “Does this task need persistent personal context, and where does that context go after retrieval?”
Default-on is not authorization, consent or availability
A default-enabled policy is an allow decision at one layer. It does not manufacture identities, grant OAuth consent, provision every application or establish commercial entitlement.
Identity and delegated access
The Work IQ API overview states that Work IQ uses delegated Microsoft Entra authentication. Requests run in the signed-in user's context; application-only authentication is not supported in that documented model.
That matters. The proposed memory feature is not evidence of a background application permission to enumerate everyone's memories. The archived notice explicitly describes only the authenticated user's profile.
Delegated access still needs careful review. A legitimate user may authorize a client whose handling of returned data is unsuitable for the organization. A user-scoped boundary limits whose context is available; it does not settle whether the recipient should process that context.
OAuth consent and tenant policy
The Work IQ permissions reference currently lists delegated permissions including WorkIQ.Read, WorkIQ.Act and WorkIQAgent.Ask, each requiring admin consent. Their descriptions distinguish reading data, acting on data and asking Work IQ agents on a user's behalf.
These are Work IQ permissions, not interchangeable Microsoft Graph scopes. Do not infer that a Graph permission such as a mail-read grant authorizes the proposed memory path. Equally, do not assume a Work IQ grant has altered an application's separate Graph permissions.
The exact permission-to-memory-path mapping was not independently verified. This article deliberately does not prescribe a memory scope, a consent command or a token request for that path.
Tenant policy then evaluates the requested operation and resource. A policy allow does not override the user's underlying permissions, and an OAuth grant does not guarantee a tenant-policy allow. Both matter.
Commercial access and actual rollout
Microsoft's API overview describes Work IQ API usage-based billing through Copilot Credits. Its memory administration page says Copilot memory is available to Copilot Chat users with and without a Microsoft Copilot license, while also marking the feature as preview.
Those statements address different products and access conditions. Neither proves that the proposed MCP memory path is available, entitled or billable in a particular tenant. Validate current prerequisites for the actual client, user and deployment cloud. A visible personalization control is not an end-to-end agent-access test.
An MCP path is not automatically a Graph endpoint
Work IQ MCP uses generic tools operating on resource paths. Microsoft's MCP overview illustrates the pattern with fetch for reading and separate tools for mutations.
The archived /memory/me/coreMemories string should be understood as a proposed Work IQ MCP resource path. It is not sufficient evidence for constructing a request to graph.microsoft.com/v1.0 or /beta with that suffix.
The current tool reference documents general Graph-backed entity tools and says its schema-discovery tools currently expose Graph v1.0 paths and schemas. It does not document this memory path. That discrepancy is another reason to wait for path-specific documentation rather than adapt an unrelated sample.
There is also a separate Microsoft Graph administrative setting: enhancedPersonalizationSetting. The resource documentation retrieved identifies it as beta and carries Microsoft's warning that beta APIs are subject to change and unsupported for production applications.
That setting controls personalization. It is not the proposed memory-read interface. A beta administrative API, a Graph v1.0 schema catalog and an MCP resource-path announcement are three different contracts. Keep their documentation and change approvals separate.
Turning memory off is not the same as deleting it
Microsoft says memory is stored in a hidden folder in the user's Exchange mailbox and benefits from mailbox security protections. However, its memory guidance explicitly says Purview retention policies and labels do not apply to Copilot memory, and there are no admin controls to enforce memory-specific retention rules.
Do not turn “stored in Exchange” into “our normal chat-retention policy deletes it.”
The same guidance distinguishes behavior by memory type:
Turning off custom instructions stops their use but does not remove them.
Turning off saved memories stops their application but does not delete existing saved memories. Deleting the originating chat does not delete those saved memories either.
Chat-history-derived details follow separate deletion behavior. Microsoft documents deletion within seven days when all originating or referencing chats are deleted, and deletion after 30 days when the relevant chat-history or Enhanced personalization control is turned off.
For user-level review, Microsoft documents Settings > Personalization. When Enhanced personalization is disabled administratively, the corresponding user controls are off and users cannot turn them on. The Graph personalization resource also documents an organization-wide setting and an optional group-based exclusion; these are not proof of per-agent MCP policy scoping.
The archived announcement says personalization settings also gate the proposed MCP profile and may take up to 48 hours to propagate. Treat that timing as reported for the planned capability. The current Work IQ governance page separately documents up to 24 hours for its policy changes. Do not substitute either interval for an emergency containment guarantee.
A practical seven-step review for IT and security teams
You can prepare without enabling an unverified path, granting additional permissions or collecting real users' memories.
Step 1: Establish the tenant-specific facts
Have the Microsoft 365 service owner check MC1539004 and roadmap item 578869 through official channels. Record the notice version, applicable cloud, deployment status and prerequisites.
Separate “announced,” “available in our tenant” and “tested successfully.” If the memory policy or path is absent, record that fact. Do not force provisioning or broaden consent merely to make an announcement's screenshots match your environment.
Step 2: Inventory clients, agents and accountable owners
For each integration that uses Work IQ, record the client application ID, publisher, business owner, agent or workload, user population, deployment location and purpose. Include shared clients used by multiple agents.
Also identify where responses are processed: the agent host, model provider, application logs, support tools and any downstream integrations. This is a recommended data-flow review, not a claim that Microsoft sends memory to all those destinations.
A named owner must be able to explain why persistent personalization improves the task. “It came with the connector” is not a business requirement.
Step 3: Review grants, not just requested permissions
Microsoft's enterprise application permissions guidance documents reviewing the selected enterprise application's Permissions, including Admin consent and User consent tabs. Use the documented administrative role appropriate to that review.
Record the resource API, granted permission, consent scope and client identity. Compare the actual grants with the workload requirement, and distinguish Work IQ grants from Microsoft Graph grants. An application's requested-permissions list alone does not prove what has been consented.
If excess grants exist, prepare a scoped remediation with the owner and test dependencies before revocation. Microsoft notes that user-consent grants cannot be revoked from that portal tab; follow its documented API or PowerShell procedure when needed. This is not a reason to run a blanket tenant-wide removal script.
Step 4: Inspect effective policy without inventing a switch
The official Work IQ governance page documents Agents > Tools, selecting Work IQ MCP in the registry, then the Policy tab when available. It describes tenant-level control; per-user, per-app and per-agent policy templates are not part of its initial control surface.
Use that verified navigation to inspect what your tenant actually exposes. The archived announcement describes a more specific memory-read policy, but do not assume the label, identifier or editability until verified in your tenant and current official guidance.
Record effective values and who can change them. Do not mistake the generic policy enablement switch for a documented memory-specific deny. Nor should you assume that disabling MCP memory access disables direct Copilot personalization or every other data-access route.
Step 5: Choose the smallest useful capability
A read-only task should not receive mutation capability merely for convenience. But read-only does not mean low sensitivity: reading persistent context may be the exact disclosure your review is meant to prevent.
Do not assume an “ask” permission is harmless because the interface accepts natural language. Microsoft's permission description for WorkIQAgent.Ask includes read and write access through accessible agents, subject to the documented authorization model.
Choose permissions and supported policy settings based on documented capabilities. Where the needed memory-specific restriction is not verified, defer memory-dependent production use rather than promise a control you have not established. Recognize the tradeoff: a broad personalization disablement can affect useful user experiences beyond one agent.
Step 6: Pilot with synthetic context and negative tests
Once official documentation and tenant availability support a pilot, use designated test accounts and non-sensitive memory content. Do not start by fetching an executive's real profile.
Validate both a permitted case and expected denials: an unapproved client, an applicable personalization disablement, and a documented memory-policy denial if that control exists. Confirm the signed-in identity and allow for documented propagation delays.
Record the tool result, not only the HTTP status. Microsoft's MCP reference warns that HTTP 200 can deliver a tool failure. If the path is unavailable or schema discovery does not expose it, that is not a successful validation. No tenant execution or sample memory response is claimed in this article.
Step 7: Set evidence, rollback and user expectations
Microsoft's memory administration guidance says memory and personalization actions do not generate Purview audit entries. Do not promise a complete memory-read investigation from Purview alone. That statement also does not establish whether the future MCP path will have separate service telemetry; its path-specific audit coverage remains unconfirmed here.
Ask the client owner what metadata is logged, how identities and requests are correlated, and how long evidence is retained. Prefer approved metadata over raw profile dumps. Establish a tested rollback for the integration and a communication plan for affected users.
Explain that “stop using memory,” “delete saved memories” and “delete chat history” are different actions. A support script that conflates them can produce false privacy assurances.
Two illustrative decisions—not observed incidents
A proposal assistant needs tone, not a personal dossier
A consulting firm wants an agent to draft proposals in each employee's preferred style. A short, explicit style preference may solve the problem without accessing a broader persistent profile.
The tradeoff is convenience versus data minimization. Automatic memory retrieval could reduce setup, but explicit task context is easier to explain and limit. Approve the smaller design first; require evidence before expanding it. This scenario illustrates a design choice, not a verified memory API response.
An MSP standardizes deployment, not consent assumptions
An MSP supports several customer tenants. One customer values personalization; another has strict restrictions on contextual information reaching third-party processing environments.
The same client package does not make their consent, policy, rollout or billing state identical. Keep a per-tenant decision record and test identity boundaries separately. Do not infer cross-customer memory access from an MSP's administrative relationship, or assume one customer's successful pilot proves availability elsewhere.
Standardizing the checklist saves effort. Standardizing on permissive defaults without the customer's decision transfers risk instead.
Key takeaways
Microsoft confirms default-on Enhanced personalization and layered Work IQ authorization. The proposed memory path, its default-enabled read policy and December rollout remain reported plans based on a secondary Message Center archive in this review.
A policy allow is not OAuth consent, commercial entitlement or proof of availability. The reported design concerns the authenticated user's profile, not every user's memory.
MCP tool paths, Work IQ permissions and Microsoft Graph endpoints are distinct. Do not invent a Graph memory endpoint or production automation from beta setting documentation.
Turning memory off is not necessarily deletion. Retention and audit limitations deserve explicit review.
Inventory clients and owners, inspect actual grants and effective policy, then pilot with synthetic data and negative tests before expanding access.
The right question is not “Can any agent read everyone's memories?” The available evidence does not support that claim. Ask instead: which approved client needs this user's persistent context, under which controls, and what can we prove about the result?
Official sources
Public Microsoft documentation checked for this draft on November 7, 2026:
Manage Copilot personalization and memory: preview status, storage, default personalization, retention, user controls and audit limitations.
Microsoft Copilot enhanced personalization control: processing purpose, confidentiality and tenant/group control.
Work IQ MCP overview: tool/path architecture and authentication.
Policy governance for Work IQ MCP: authorization layers, tenant-level policy, documented admin navigation and propagation.
Work IQ API permissions reference: delegated permissions and admin consent requirements.
Microsoft Work IQ API: delegated-only authentication and usage-based billing model.
Work IQ MCP tool reference: tools, schema-discovery limitations and tool-result errors.
Enhanced personalization setting, Microsoft Graph beta: administrative resource and beta warning.
Review permissions granted to enterprise applications: grant review and revocation procedures.
Microsoft 365 roadmap, requested item 578869: official verification destination; item-specific details were not retrievable in this review. This link is not cited as confirmation of the reported dates.
The secondary Message Center archive linked earlier is identified as such and is not an official Microsoft source.
About IAMPathway
IAMPathway publishes practical guides on Identity & Access Management: Microsoft Entra ID, governance, identity threats, Application Governance and AI Governance. Browse all topics or read the glossary.