Personalized AI infrastructure
Superintelligence Data Export Test: Can You Take Your Memory Out?
A short export test to check whether your prompts, agent memory and logs can leave your AI vendor in usable formats before you sign or expand.

The superintelligence data export test is a short check that asks whether your company can take its prompts, agent memory, tool settings, permissions and activity logs out of an AI vendor's system and read them in usable formats without that vendor's help. If the answer is no, the agent is a rental that holds your operating knowledge, not infrastructure you control. Run the test before you sign, expand or renew any contract that stores company memory.
Key takeaways
- Exportability is a property of the setup, so test it before the contract, not after the first dispute or incident.
- Ask for five things: prompts, memory, tool and permission configuration, and logs, each in a format a person or another system can read.
- A complete export with unlabeled memory is nearly as hard to use as no export, so check for source, owner and date on every memory item.
- Run a restore check in a separate workspace. A file that exists has not yet proven that the role can be rebuilt from it.
- Keep the inventory and the export routine under your own control, and repeat the test on a schedule.
What the export test actually checks
Most leaders hear a vendor say that the data belongs to them and assume the question is settled. Ownership on paper and usability in practice are different things. A vendor can be accurate about ownership while the only way to see an agent's memory is through its interface, in a proprietary format, with no timestamps and no record of which tool calls led to which actions.
The test checks three things. First, coverage: is everything the agent needs to behave the way it does included in the export? Second, legibility: can someone who was not in the original project read the material and understand it? Third, portability: can a second system or a second team use the material to rebuild the role somewhere else? The first two are inexpensive to check. The third is the one most companies skip, and it is the one that decides whether you are locked in.
Treat the test as an operations and procurement question, not as a technical audit reserved for engineers. The business owner needs a clear result: a yes, a no, or a list of gaps with named owners and a deadline for closing each one.
Build an inventory before you ask for anything
You cannot test an export you have not defined. Before you contact a vendor, write down what the agent's operating context consists of. A usable inventory has five parts.
- Instructions: system prompts, role descriptions, and the stop rules that tell the agent when to hand work to a person.
- Memory: the facts the agent retains about customers, suppliers, policies and past decisions, plus where each fact came from.
- Tools and connections: each system the agent can read or write, the account it uses, and what it is allowed to do there.
- Permissions: who approved each capability, the threshold above which a person must confirm an action, and when that was last reviewed.
- Logs and evaluations: the record of what the agent did, the test cases used to judge its quality, and the results of those tests over time.
If any of these five parts lives only in someone's head or in a chat thread nobody can find, that is your first finding. For a practical way to draw these components on one page, see Draw Your AI Stack: A One-Page Architecture Map for Leaders.
Run the test in five steps
The test itself takes an afternoon for one agent role. Use the same steps for each vendor and each agent so the results can be compared over time.
- Step one: Ask the vendor, in writing, for a full export of one agent's prompts, memory, tool configuration, permission settings and logs for a defined period. Record the date of the request.
- Step two: Check the files against your inventory. Mark every inventory item as present, partial or missing. Missing items count as findings, not as small details.
- Step three: Open each file without the vendor's interface. Plain text, Markdown, CSV and JSON can be read with common tools. A proprietary binary, a screenshot or a link that expires is a failure of legibility.
- Step four: Check that memory entries carry a source, an owner and a date of last confirmation. Check that logs carry timestamps and identify the tool calls and any human approvals involved. Without these fields, the memory is a pile of statements, not a record.
- Step five: Restore the role in a separate workspace, using only the exported material and a written setup guide. Ask a colleague who did not build the agent to run three realistic requests and compare the outputs with what the original produced. Record every place where they had to guess.
Score the result as pass, pass with actions, or fail. A pass means all five steps are complete and the restore produced equivalent behavior on the test requests. A pass with actions means the gaps are small, named and assigned. A fail means the role cannot be moved in its current form, and you should weigh that before expanding its scope.
Illustrative scenario: A ten-person distribution company asks its vendor for an export of its order-status agent. The prompts arrive as plain text and the logs arrive as CSV, which is good. The memory arrives as one JSON file with customer notes but no source or date fields, and the permission settings exist only as a screenshot of an admin page. The restore test in a separate workspace fails on two of three requests because the rebuilt agent cannot tell which discount rules are current. The company records a fail with actions, assigns the memory fields and the permission export to the vendor, and schedules a repeat test for the following quarter.
Judge the formats by what a stranger could use
Format matters less than completeness and labeling, but some formats make the work easier. Prefer files that a standard tool can open and that carry structure: JSON for records with fields, CSV for logs and tables, Markdown or plain text for instructions. Avoid exports that depend on the vendor's viewer to make sense.
Labels matter more. Every memory item should answer three questions: where did this come from, who is responsible for keeping it correct, and when was it last confirmed. Logs should let you follow one request from the first instruction through each tool call to the final action and any human approval. The concepts behind this kind of record, including logs, traces and metrics, are explained in plainly written form in the [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/). You do not need to adopt the technology to borrow the questions it asks.
Check the boundary between memory and model as well. Some vendors store knowledge inside the model through fine-tuning, which is hard to inspect and harder to move. Retrieval keeps knowledge in documents and records you can read and replace. The [AWS guidance comparing retrieval and fine-tuning](https://docs.aws.amazon.com/prescriptive-guidance/latest/retrieval-augmented-generation-options/rag-vs-fine-tuning.html) sets out the tradeoffs. For a company that wants to keep control of its knowledge, retrieval over records it holds is usually the easier material to take with it.
What a failed test tells you
A failed export is information, not a verdict on the vendor's character. It shows where your operating context actually lives, which is often more revealing than the sales material. Common findings include memory that exists only inside a chat history, permissions set by one person who has since left, and logs kept for a short window and then deleted.
Each finding should lead to a decision. Either the vendor closes the gap within a stated period, you move the material into storage you control, or you narrow the role until the gap no longer matters. Do not expand the agent's scope while a finding is open. The same discipline applies to the approval limits the agent operates under. If a limit needs to change, route the change through a documented process, as described in How to Request and Approve Changes to a Managed AI Agent.
Be wary of answers such as we can build that for you later, given with no written scope. A useful reply names the artifact, the format, the owner and the delivery point. Anything vaguer suggests that the export depends on the relationship rather than on the system, which is exactly the lock-in the test is meant to expose.
How Be Human handles exportability
Be Human designs and operates AI agents so that their identity, memory, channels, tools, permissions and operating context stay under the client's control. The first engagement defines one role, the memory and tools that role needs, its approval boundaries and an observable success criterion. Because the inventory is part of that first definition, the company knows from the start what should be exportable, rather than discovering the gaps during a dispute. Managed operations then keep the inventory current through monitoring, permission reviews, recurring evaluations and controlled changes, so the export test can be repeated on a schedule instead of reconstructed in a crisis. This describes the service approach, not a promise about any particular platform.
Start with one role and test it this quarter
Pick one agent role, ideally the one that touches the fewest systems and carries the least risk, and run the five steps on it before testing anything larger. Write down the inventory, request the export, and record the results where the operations lead and the role owner can both see them. Once the first pass is complete, repeat it for each additional role and each vendor you use, and compare the results. If you are unsure which role should go first, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Does passing this test mean we meet privacy law? No. The test shows whether you can move and read your operating material, which is a precondition for managing it well, but it does not by itself establish lawful processing of personal data. For governance questions, the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) offers a structured vocabulary for mapping risks, and your legal or compliance advisers should confirm the obligations that apply in your jurisdiction.
Can we run the test without the vendor's cooperation? Only partly. You can always inventory what you have configured, check what you already hold in your own systems, and document where memory is created. A full export, however, requires the vendor to provide it, which is why the request should be written, dated and tied to a defined period and a defined set of artifacts.
How often should we repeat the test? Repeat it whenever a role changes materially, such as a new tool connection, a new data source or a change in who approves actions, and also on a regular schedule. Many teams pair it with a quarterly permission review so the evidence stays current. A test that was passed a year ago and never repeated tells you very little about the agent you run today.
Does retrieval-based memory make exports easier? Usually, yes, because the knowledge sits in documents, records and indexes that you can inspect and replace. It does not remove the need for labels, owners and logs, since a well-organized folder of stale or unattributed facts is still hard to trust. Treat retrieval as a way to keep your memory in forms you control, then apply the same labeling and restore checks described above.
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.


