Digital twins used by AI agents need more than a live snapshot: they need a dependable history of what was known, when it applied, where it came from, and why a recommendation changed. Without that context, an agent can see the current sensor reading but miss the maintenance event, operator note, or earlier assumption that gives it meaning.
Why a digital twin needs history, not just current state
A digital twin can represent an asset, process, or system as it exists now. But an agent supporting decisions over time also needs to understand how that state developed. A temperature reading, for example, is more useful when considered alongside the equipment relationship it belongs to, a maintenance record, an operator’s note, and prior recommendations.
As an Amazon Associate I earn from qualifying purchases.
When those details live in separate systems, the agent must repeatedly retrieve and reconcile them. That makes it harder to preserve context from one interaction to the next and for an operator to understand what informed a decision. In an InfoWorld opinion article published August 27, 2026, Tobie Morgan Hitchcock argues that digital twins are becoming decision environments, not merely virtual counterparts. That is an architectural perspective, not a description that applies to every deployment.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →What an AI agent should remember
Useful memory is not simply a larger store of text. It should preserve enough context to make operational facts interpretable and recommendations inspectable.
#1 Best Overall
- Source: which sensor, system, document, person, or prior process supplied the information.
- Time: when the information was recorded, when the agent learned or believed it, and the period in which the underlying condition applied.
- Uncertainty: how reliable or tentative the information is, rather than presenting every assertion as equally certain.
- Relationships: how assets, components, events, constraints, and records connect to one another.
- History and disagreement: what changed, what superseded an earlier fact, and which claims conflict.
- Reasoning context: the documents, relationships, previous interventions, and assumptions that contributed to a recommendation.
Hitchcock summarizes his position this way: “The more durable approach is to treat agent memory as first-class twin data.” This is his recommendation as an InfoWorld contributor and CEO and co-founder of SurrealDB, not a standard requirement or a proven universal design.
How to tell when a fact was true
Operational systems often blur distinct kinds of time. A reliable history should not treat a timestamp as if it answered every temporal question.
Rank #2
- Valid time: when a condition was true in the physical or operational world. A component may have been failing before a sensor alert was entered.
- Recorded or learned time: when a system recorded the information or the agent became aware of it.
- Belief time: the interval during which the agent treated a claim as current or credible.
These times can differ. A maintenance report entered after a repair might describe a condition that existed earlier; a later correction may change the agent’s belief without changing the historical period in which the original condition applied. Retaining those distinctions lets an operator reconstruct what the agent could have known at a particular point and why its view later changed.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →How operators can inspect a changed recommendation
A changed recommendation should be traceable to the information that changed, not left as an unexplained difference between two outputs. A useful record connects the recommendation to relevant retrieved documents, asset relationships, earlier interventions, and assumptions. It should also show which facts were revised or superseded and when.
That trace helps answer practical questions: Did a new reading alter the assessment? Was an older maintenance note corrected? Did the agent rely on an assumption that no longer holds? Keeping the evidence and its history together makes these questions easier to investigate than reconstructing the decision from disconnected logs.
What architecture choices involve
Hitchcock contrasts prompt-window stuffing, separate memory stores, and treating memory as part of the twin’s data. A prompt can provide context for a particular interaction, but it is not by itself a durable, governed operational history. Separate vector, key-value, graph, or document stores can each serve useful purposes; they can also introduce synchronization, access-control, and governance seams when their contents must remain consistent.
His preferred direction is a shared data substrate able to represent structured, graph, document, vector, and temporal information under common governance and transaction rules. That is a design recommendation, not evidence that one database approach will outperform others in every workload. Teams should compare architectures against their own requirements:
- Provenance, auditability, and the ability to preserve conflicting claims.
- Support for valid-time and belief-time history.
- Query and retrieval patterns, including how recommendations trace back to evidence.
- Consistency requirements and the boundaries across which transactions must be reliable.
- Access control and whether memory is local to an agent, shared among agents, or both.
- Integration and synchronization work across existing systems.
- Latency, scale, and operational ownership.
The cited argument is qualitative; it does not provide measured comparisons of vendors or architectural patterns. A unified substrate may simplify some coordination, while a distributed design may fit existing systems or workload boundaries better. The choice depends on the constraints the team must meet.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What ISO 23247 does—and does not—cover
ISO 23247 provides framework context for manufacturing digital twins. Part 1:2021 sets out an overview and general principles, while Part 2:2021 provides a reference architecture. Part 5:2026 addresses a digital thread for creating, connecting, managing, and maintaining manufacturing twins across lifecycle stages. Part 6:2026 addresses composition, including integrated, unified, and federated approaches to interoperability.
These standards are relevant to lifecycle and system-of-systems design, but they do not prescribe persistent AI-agent memory or endorse a single database substrate. They provide context for architecture and interoperability, not validation of the performance claims in an individual opinion article.
How strong is the case for digital twins?
InfoWorld reports that 62% of surveyed C-suite executives said they got immense value from digital twins, attributing the figure to a 2024 Hexagon survey. This is secondary reporting; the figure should not be treated as independently verified from Hexagon’s original survey. It also does not establish that a particular memory architecture produces better outcomes.
The practical case for memory is narrower and clearer: when agents make or support decisions across repeated interactions, preserving provenance, timing, relationships, and changes can make their context more coherent and their recommendations easier to inspect. Whether to centralize that data or connect multiple stores is an architecture decision to evaluate against the workload, governance needs, and operational constraints.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
