Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

Superintelligence Layer for Multi-Site Companies: One Memory, Many Sites

A practical model for sharing one AI memory and permission set across company locations while keeping each site's rules and customer data separate.

Be Human typographic cover with the title: Superintelligence Layer for Multi-Site Companies: One Memory, Many Sites
Conceptual illustration created with AI · Be Human

For a company with several locations, AI works best when one company-level memory holds what every site must share, such as the service standard, product definitions and approved wording, while each location keeps its own permissions, customer records and local rules. The model is simple to describe and easy to get wrong, because a shared memory without site boundaries lets one branch's exception, price or customer history appear in another branch's answer. This article explains how to separate shared from local, where to enforce that separation, and how to prove it holds.

Key takeaways

  • Sort every data source into shared, site-scoped, customer-scoped or restricted before connecting an agent.
  • Enforce site boundaries in the data and tool layer, not only in the instructions given to the model.
  • Treat cross-site access as an explicit exception with an owner, a reason, an expiry date and a log entry.
  • Test isolation with deliberate cross-site questions, not only with normal daily use.
  • Review scopes on a fixed schedule, and again whenever a site opens, closes or changes its manager.

Why multi-site companies need one memory and several boundaries

Multi-site businesses share a repeating problem. Each location answers the same questions about returns, booking rules, coverage or onboarding, and each answer drifts a little over time. A company-level memory corrects that drift by giving every agent and every person one current version of the service standard. That is the useful half of the design.

The risk sits in the other half. Memory is not one thing. In practice it has three layers. Company memory holds policies, catalog facts, brand voice and escalation paths. Site memory holds opening hours, the local staff roster, neighborhood partnerships and local supplier terms. Customer memory holds the history of individual people, which usually belongs to the location that served them. An agent that reads all three without distinction will eventually answer a question about one site using a fact from another.

Risk frameworks such as the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) ask organizations to map where AI systems touch data and people, and to assign responsibility for each touchpoint. For a multi-site company, that mapping is the first practical step. It turns a vague worry about AI and data into a short list of specific flows that you can approve or block. For the wider picture of how these pieces fit, read the seven layers of company AI infrastructure.

Separate what is shared from what is local

Start with a classification, not a technology choice. Every source the agent may read receives one of four classes, and each class carries a default visibility.

  • Shared: service standard, product definitions, approved wording, brand voice and escalation policy. Every site reads these by default.
  • Site-scoped: opening hours, local staff roster, local supplier contacts and site-specific promotions. Only that site's agent reads these.
  • Customer-scoped: appointments, complaints, purchase history and anything containing personal data. Visible to the location that holds the relationship, and to others only under a written rule.
  • Restricted: site financial results, contracts, legal matters and personnel files. Kept outside the agent unless a named role needs them for a defined task.

Each class needs one named owner who can change its status. Without an owner, a site manager will quietly add a local price list to the shared memory because it is convenient, and the boundary disappears without anyone deciding to remove it. Keep the classification short enough that a manager can recite it from memory. A narrower version of this design for two locations, with a shared memory and separate permissions, is covered in the two-location guide.

How do you enforce site boundaries in the tools?

The most common mistake is to place the boundary in the prompt, with a line such as 'only answer about the Northside site.' That is a request, not a control. Anyone who can phrase a message cleverly may try to override it, and prompt injection is a known class of attack against language model applications. The [OWASP guidance on prompt injection prevention](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) makes the same point in practice: instructions alone should not be the only defense. The control belongs where data is fetched.

Retrieval should filter by site identifier before any content reaches the model. A record that carries a site tag is returned only to a request authorized for that site. The authorization decision belongs in the tool that queries the data, based on the identity of the person or service making the request, never on a value the model supplied. The [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) describes this pattern: check access on every request, at the point where the resource is read, and deny by default.

Apply the narrowest scope that completes the task. A location assistant that books appointments needs write access to its own site's calendar and read access to the company service standard. It does not need another site's customer list, even if the same software platform hosts both. Separate credentials for each tool and each site make the boundary auditable, and the agent should never be able to widen its own scope. The permission review process is described in how to run a permission review for your AI agents.

A step-by-step plan for the shared layer

The following sequence is a practical starting order for a company with several sites. It is written for an operations leader who will sign off each step.

  • Inventory the sources. List every system and document the agent may read, with its owner.
  • Classify each source as shared, site-scoped, customer-scoped or restricted, and record the decision.
  • Tag every site-scoped and customer-scoped record with a site identifier at the source, not later in the pipeline.
  • Write the company memory once, with a named editor and a change log, and let every site read the same version.
  • Create one credential per tool and per site, limited to the minimum read or write actions the role requires.
  • Define the cross-site exception: who may approve it, what reason is required, how long it lasts and where the access is logged.
  • Build a boundary test set of questions designed to tempt the agent across sites, and run it before any site goes live.
  • Set the first review date and the evidence each review must produce, such as the scope list, the exception log and the test results.

Illustrative scenario: two sites, one assistant

Illustrative scenario, not a real client: a company runs two service locations in different cities and shares one booking assistant. The assistant reads the company service standard, plus each site's hours and staff roster. The Northside site offers a seasonal package for a local school. The Riverside site has never offered it.

A Riverside customer asks whether the school package is available. A well-built assistant answers from shared policy: packages are offered by individual locations, and Riverside can confirm its own options. It does not reveal the Northside school contract, the contact name or the customer list, because those are site-scoped or customer-scoped records that the Riverside credential cannot read. The assistant can offer to arrange a call with the Riverside team.

The test comes next. A Riverside staff member, acting as a tester, asks the assistant to list recent Northside bookings for comparison. The expected result is a refusal that explains the request is outside the site's access, logged with the requester and the blocked action. If any Northside detail appears in the answer, the boundary has failed, and the change goes back through review before the agent is used again.

In a managed engagement, Be Human starts by writing down the role, the memory each site needs, the tools it uses and the permissions each tool receives. The classification and site identifiers are implemented in the data and tool layer, and the boundary test set runs before launch. After that, monitoring records which sources each answer used, scope changes go through a documented request that the owner approves, and permission reviews produce evidence on a schedule. The goal is that every answer can be traced to a site and a permission, rather than relying on the model to behave.

How do you test isolation and keep it true over time?

A boundary tested once is still a hypothesis. Tests should include obvious cross-site requests, indirect ones (for example, an average wait time across all locations, which can reveal one site's pattern), and requests framed as part of ordinary work. Keep the test set alongside the agent configuration, and rerun it after every change to the memory, the tools or the underlying model.

Logging makes the boundary visible after launch. Each answer should record which site was in scope, which sources were retrieved and which tool calls were made. The [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/) describes the general idea of traces and telemetry for understanding complex systems. Applied here, it means you can reconstruct, after an incident, which record reached which answer. Keep those logs for a defined period, under the same access rules as the data they describe.

Boundaries drift as companies change. A site closes, a manager moves to another location, a partner joins, or a new service line appears. Each event is a trigger for a scope review, and each change should pass through a written request so the owner sees the boundary change, not only the new feature. The process is described in how to request and approve changes to a managed AI agent. If you are not sure which records in your locations should stay local, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

Can one AI agent serve several locations without mixing their data? Yes, if the boundary is enforced where records are fetched rather than requested in instructions. Each record carries a site identifier, each credential is scoped to one site and task, and cross-site access is treated as a logged exception. Test the boundary with deliberate cross-site requests before launch.

Should customer records be shared across locations? Only under a written rule that the company has approved. By default, a customer's history belongs to the location that holds the relationship. A shared customer profile can be justified for a single program, but it needs its own owner, stated purpose and review date.

How many memory layers does a multi-site company need? Three working layers are enough for most multi-site setups: company memory, site memory and customer memory. Restricted financial and legal material usually stays outside the agent altogether. The exact split matters less than naming an owner for each layer and recording who may change it.

Who should approve a cross-site exception? The owner of the data should approve it, and the operations leader should be accountable for the exception record. The approval should state the reason, the scope, the expiry date and the evidence that access ended. Exceptions without an end date tend to become permanent without anyone deciding that they should.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →