AI Agent Memory Scopes: Account, Team and Company-Wide Rules
Set three AI agent memory scopes, account, team and company-wide, and the rules for which facts may move up a level, with named owners, checks and an audit trail.

An AI agent memory scope is the set of people and records that can read a stored fact, along with the rules for who may add or change it. Most business agents need three scopes: account memory for one customer, asset or job, team memory for the working knowledge of one group, and company-wide memory for approved policies and procedures. A fact should move up a level only when it holds beyond the record it came from, a named owner has approved it, and it carries a review date.
Key takeaways
- Use three scopes: account, team and company-wide. Give each one an owner, a clear purpose and a retention rule.
- A fact moves up only when it applies beyond its original record and an accountable person approves the change.
- Staff opinions, one-off requests and unverified summaries stay at account or team level, labeled as what they are.
- Every promotion leaves a log entry with the source, the approver, the date and the previous version, so it can be reversed.
- Start with one role and one boundary, then widen scopes after a review rather than before the first agent goes live.
What the three memory scopes mean
Account scope holds facts tied to one customer, property, vehicle, project or supplier relationship. An agent should load these facts only when it is working on that record, and the people who may read them should match the people who already see that record in your systems. A preferred delivery window, an open warranty claim or a note that a customer asked not to receive phone calls all belong here.
Team scope holds working knowledge shared by a group that does the same kind of work. A dispatch team might share how it sequences routes on busy mornings, and a billing team might share which invoice fields usually need a second look. Team memory helps an agent behave like a member of that team, but it should not spill into work done by people outside the team. Owners usually sit inside the team, and changes should go through the team lead.
Company-wide scope holds facts that apply to everyone: approved policies, the current version of a procedure, product rules, service boundaries and approved wording for common disclosures. It is the most useful scope and the most dangerous one to fill carelessly, because an error there reaches every role at once. Keep it small, versioned and reviewed on a schedule. Anthropic's guidance on context engineering makes a related point: an agent works best when it receives the smallest set of relevant information for the task in front of it. For the boundary between personal and shared memory, see Shared vs Personal Memory in Company AI Agents.
Which facts may move up a level
Promotion means elevating a fact from a narrower scope into a broader one. Treat it as a change to company policy, not as a convenience. A fact may move up only when it passes every check on this list:
- It holds for more than one account or team, and you can name at least two records where it applies.
- It comes from an approved source, such as a signed policy, a current procedure document or a verified record, not from a single conversation.
- It contains no personal data that the broader audience does not need. Keep names, contact details and health details at account level.
- A named owner has approved it, and that owner is accountable for the process it describes.
- It has a review date, after which the owner must confirm it, correct it or retire it.
- The agent can cite where the fact came from, so staff can check it quickly.
If any check fails, the fact stays where it is. That rule is deliberately conservative. Missing a promotion costs a little repetition, while a wrong promotion can give the wrong answer to every customer at once. When two promoted facts disagree, When AI Agent Memories Conflict explains how to decide which one wins. The NIST AI Risk Management Framework is a useful reference for assigning accountable owners to AI systems in the same way.
Illustrative scenario: a repair note that should not go company-wide
This is an illustrative scenario, not a description of a real client. A small property management firm uses an agent to answer tenant messages and prepare maintenance tickets. A tenant reports a slow drain in one unit. The drain report belongs in account scope for that unit, because it describes one property and one request. A technician also writes that one subcontractor is always late. That is a staff opinion, so it belongs in team scope with a label showing who wrote it and when, and the agent should never repeat it to tenants.
Later, the office notices the same drain problem keeps returning in buildings that share one type of fitting. The firm writes a short procedure: check the fittings on a named list before sending a ticket for a slow drain. The owner reviews the procedure, confirms its source, sets a review date and promotes the procedure, not the individual complaints, to company-wide scope. The complaints stay where they were, and the audit log shows exactly what moved and why.
Permissions, write rules and the audit trail
Scopes are only as strong as the permissions behind them. Each scope should have a list of roles allowed to read it and a shorter list allowed to write to it. An agent acting for a dispatcher should not be able to read a billing team's notes simply because both sit in the same system. The OWASP authorization guidance describes the general principle: check access where the data is used, and deny by default. Apply that to memory the same way you would apply it to a customer record. For more on access by role, see Role-Based Access to AI Agent Memory.
Write rules need the same care. Decide who may add a fact to each scope, who may correct it and who may remove it. Keep an audit trail that records the source, the approver, the date and the earlier version. When a fact turns out to be wrong, you can then find every place it was promoted and fix them together rather than hunting through conversations one at a time.
How Be Human handles memory scopes in managed operations
Be Human defines scopes as part of the first engagement, alongside the one role the agent will perform. We list the facts each scope may hold, name the owners who approve changes, set the promotion checks described above and configure the agent to read only the scopes its role needs. In managed operations, we review pending promotions on a regular schedule, run recurring evaluations that test whether the agent answers from the correct scope, and revisit permissions when staff change responsibilities. The aim is a boundary your team can inspect and adjust, not a promise that the agent will never make a mistake.
Where to start: a plan for your first scope boundary
Start with a narrow boundary: one role, one team and a short list of approved company documents. Work through these steps in order:
- Name the role the agent will perform and list the records it touches, such as tickets, quotes or customer files.
- Write down the facts that role needs, then sort each one into account, team or company-wide scope.
- Assign one owner to each team and company-wide fact, and set a review date for each.
- Apply the promotion checks to any fact you want to move up, and log the decision with its source.
- Test the agent with realistic questions drawn from recent work, and confirm that it answers from the right scope and cites its source.
- Widen the scopes only after a review shows the first boundary holds.
If you are not sure which facts in your business belong at which level, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Can one memory store serve the whole company? A single store is technically possible, but it removes the boundaries that keep customer details away from staff who do not need them. Scopes are a labeling and permission layer on top of storage, so you can apply them even when the underlying technology is shared. Decide first what each scope may hold, then choose how to store it.
Who decides when a fact moves up a level? The owner of the process the fact describes should approve the promotion, and the approval should be recorded in the log. Operations leadership sets the rules and settles disagreements, but a fact should not move up simply because the agent or one staff member wants it to. A short review of pending promotions on a regular schedule is usually enough.
What happens to account facts when a customer leaves? Account facts should follow the retention policy you already apply to the customer record, rather than staying in memory indefinitely. Decide whether each fact should be corrected, expired or removed, and record which action was taken. If you need to remove stored content, How to Delete AI Agent Memory walks through the options.
Do we need all three scopes from the first day? Usually not. A first engagement often needs account memory for the records the role touches and a small team scope, with company-wide memory added only when there is an approved document set worth sharing. Adding a scope later is easier than removing a wrong fact after it has reached everyone.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


