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

Context & memory

Superintelligence Memory Model: Entity Records That Outlast Any Model

An AI agent data model built on stable entity records for customers, jobs and assets, so company memory survives a model change and stays under human control.

Be Human typographic cover with the title: Superintelligence Memory Model: Entity Records That Outlast Any Model
Conceptual illustration created with AI · Be Human

An AI agent data model is the set of records an agent reads and writes to keep company memory: stable entity records for customers, jobs and assets, each with a fixed identifier, defined fields, a named owner and a source for every fact. Because these records live in storage the company controls rather than inside a model's weights or a chat log, a new model can read them on its first day of use. This article explains how to define those records and what to decide before an agent touches them.

Key takeaways

  • Store company memory as entity records with stable identifiers, not as chat transcripts or long prompt text.
  • Start with three entity types, customers, jobs and assets, and give every field one owner role and one source.
  • Keep confirmed facts, staff notes and agent-written summaries apart so a reader can tell what is verified.
  • Version the schema and log every change, so that a model swap cannot quietly change what a field means.
  • Test recall against known questions after each model or prompt change, before the agent gets wider access.

Why company memory should not live inside the model

Models get replaced. A provider retires a version, a better model appears, or a team decides to route one workflow to a different tool. If what the company knows about its customers exists only in conversation history or inside a prompt, every change means rebuilding that knowledge by hand, or losing part of it. Context windows add a second limit. An agent can only reason over what is loaded into a single request, so older facts drop out unless something outside the model selects and carries them forward.

The practical answer is a record layer the company owns, with the model acting as a reader and writer of that layer. The layer is not a copy of every document. It is the agreed version of the facts the business relies on: who the customer is, what work was done, which equipment is installed and what is still open.

This does not make conversation history useless. A summary of a support thread helps with the next message. The difference is authority. A summary is a convenience for one interaction, while an entity record is the reference that other interactions depend on. The distinction between the two is explored in AI memory vs chat history, and the loading side of the problem is covered in context window explained for business.

What an entity record contains

An entity record describes one thing the business tracks over time. A good record is plain on purpose. Every customer, job and asset has the same basic shape, so a model reading the record does not have to guess where a phone number, a service date or a serial number is stored. A workable minimum includes:

  • A stable identifier, assigned once and never reused for a different thing
  • An entity type drawn from a short, fixed list
  • Core attributes, each with a defined format such as text, date or enumerated status
  • Relationships to other entities, stored by identifier rather than by name
  • A source reference that says where each fact came from
  • A status such as active, paused, closed or archived
  • An accountable owner role responsible for correctness
  • A change log that records what changed, who approved it and why

Names are the weakest key a memory system can use. Two people share a name, a company changes its trading name, and one street address is typed three different ways. Identifiers stop a model from merging two customers because their text looks alike. The rules for deciding when two records describe the same party are a separate design task, covered in customer identity matching for an AI agent.

Customers, jobs and assets: why these three

Most service businesses can organize their memory around three entity types. A customer is the party who requests or pays for work. A job is a unit of work with a scope, a date window and an outcome. An asset is something that persists across jobs: a vehicle, a machine, a building system, a rental unit or a piece of installed equipment. The relationships are simple. A customer requests jobs, a job touches one or more assets, and an asset's history is the set of jobs that served it.

This structure answers the questions staff ask most often. What was done here before? What is still open? Which equipment is covered, and by which visit? When those answers come from records instead of re-derived text, the agent is faster and easier to check. Adding more types early tends to make the schema harder to keep correct, because each new type brings its own owners, statuses and update rules. Begin with the three that your first workflow actually touches, and add others only when a real question cannot be answered without them.

A step-by-step plan for the first schema

Defining the schema is an operational exercise, not a database exercise. The following sequence keeps it tied to work the business needs done:

  • Step 1: Choose the one workflow the agent will support first, and write down the questions it must answer correctly.
  • Step 2: For each question, list the fields needed to answer it. Remove every field that no question uses.
  • Step 3: Write the three entity types with those fields, one line per field, including format and allowed values.
  • Step 4: Name one owner role for each field and the system where that field is authoritative.
  • Step 5: Mark each field as a confirmed fact, a staff note or an agent-generated summary.
  • Step 6: Define the identifier format, the rule for merging duplicates and the role allowed to approve a merge.
  • Step 7: Decide which fields the agent may write directly, which it may only propose, and which need human approval.
  • Step 8: Test recall with a set of questions whose correct answers your staff already know, and record every miss.

Keep the schema in version control or in a dated changelog. Each change should state which questions it affects, so that a later reviewer can see why a field exists and what breaks if it is removed. The permission side of step seven should follow explicit rules per role and per action rather than a broad grant. The OWASP Authorization Cheat Sheet is a useful reference for that discipline. The schema also needs an intake path for new entities, which is where the entity memory schema for superintelligent agents gives more detail on field design.

Keeping confirmed facts apart from opinions and drafts

A record that mixes what a customer said, what a technician suspected and what an agent inferred will mislead the next reader, human or machine. Each fact should carry its source type. A confirmed fact comes from an invoice, a signed document, a system event or a correction approved by the owner role. A staff note is useful but carries the author's judgment, so it should be labeled as such. An agent summary is a draft until someone checks it against the records.

The practical rule is that the agent answers from confirmed facts and presents staff notes as notes. When it must rely on a draft, it should say so. This is the main reason to store the source type as a field rather than as a comment in free text. The reasoning behind this split is explored in customer facts vs staff opinions in AI agent memory. The NIST AI Risk Management Framework offers a broader vocabulary for governing systems like this, including how to document what a system relies on and who is accountable for it.

Illustrative scenario: a repair company that changes models

Illustrative scenario, not a real client. A residential appliance repair company runs an agent that answers schedule and warranty questions. Its records hold a customer identifier, an asset record for each installed water heater with the serial number copied from the invoice, and a job history linked to those assets by identifier. Later, the company moves the agent to a newer model. Because the agent reads the same records, its answers about warranty dates and past visits stay consistent with what staff already see.

Some time after the move, a dispatcher notices that one customer's service address is missing a unit number. The correction is made on the customer record by the owner role for addresses, the change log shows who approved it, and the next answer reflects the fix. Nobody had to re-explain the company's equipment to a new system, and nobody had to search old chats for the correct address. That is the practical value of a record layer that outlasts the model.

How Be Human approaches this in a managed engagement

Be Human designs and operates this record layer as part of a first engagement, which defines one useful role, the memory and tools that role needs, its approval boundaries and an observable success criterion. The entity schema is written with the client's named owners, versioned, and tied to the questions the role must answer. After launch, managed operations cover monitoring, permission review, integration maintenance and recurring evaluations, so the records and the agent's access stay current as the business changes. When a model is replaced, recall is tested against the existing records before the new model handles production work. Be Human works remotely, in English and Portuguese, and the first step is usually a short review of where customer, job and asset facts live today. If you are not sure which task in your business should be the first one, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Do I need a database to hold entity records? Not at the start. A structured store, such as a relational database or a set of tables in a system you already run, is enough for customers, jobs and assets in most small operations. What matters more than the technology is that every record has an identifier, an owner and a change log, which a shared spreadsheet rarely enforces once several people edit it.

Can a vector database replace entity records? No. Vector search finds passages that resemble a question, but it does not confirm that the matched text is the current fact about a particular customer or asset. Entity records give the agent exact lookup by identifier, and retrieval can then supply the supporting documents around those records. The two work together rather than substituting for each other.

How do entity records help when we switch AI models? The records sit outside the model, so a new model reads the same identifiers, fields and history. What changes is how well the model reads and writes them, and that is what you test against known questions before rollout. Because schema changes are logged, a change in meaning does not pass unnoticed.

Who should own the records? Each field should have a named owner role, usually the person accountable for the process that produces the fact, such as billing for invoice data or dispatch for scheduled visits. The agent can propose changes, but approval and the final write belong to the owner role unless you have explicitly delegated that write. Review ownership whenever responsibilities move between people.

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 →