Architecture
The infrastructure.
Piece by piece.
UBcore OS is not an interface over someone else's models. It is a proprietary architecture designed for environments where data matters and errors cost. This page shows its architecture, component by component.
Where UBcore lives
The plane where the company decides.
At the base of a company's technology stack are the transactional systems: management systems, sales systems, document systems, collaboration infrastructures. They keep operations running. They are where the company records what it does.
Above them live the operational automations: workflows, specialized agents, integrations, execution bots. They build speed on top of those systems. They are where the company accelerates what it does.
Above both there is a space that today the products on the market do not inhabit: the Cognitive Layer. It is where the company decides what is worth doing, on which principles, with which criteria, in coherence with which history. This is where UBcore lives.
The Cognitive Layer does not replace what sits below it. It integrates it from above: it knows what the transactional systems record, knows what the operational automations are doing, and on top of those two flows it builds the reasoning that makes it possible to choose.
The two worlds do not exclude each other. They complete each other.
The flow
From input to answer.
What happens when a user asks a question. Every request travels through five steps before generating an answer. None of them is optional.
AUTHENTICATION
The user logs in with their own credentials. The system verifies identity and permissions. Each client sees exclusively their own data: the separation is structural.
STRATEGIC CONTEXT
Before formulating the answer, the system loads the client's strategic context: the company constitution, the non-negotiable truths, the criteria of choice. This context enters every response with priority over operational documentation.
RETRIEVAL
The question is matched against the client's document base. Knowledge is organized on an internal hierarchy: strategic truths take priority over operational documents, which take priority over the context of projects in progress. Selection follows this hierarchy, not just textual similarity.
INFERENCE
The question together with the retrieved context is sent to the reasoning model chosen by the client. The choice belongs to the client: commercial frontier models for cases that demand the highest quality, local on-premise models for contexts where data cannot leave the corporate perimeter. The system is agnostic to the engine.
ANSWER WITH SOURCES
The user receives the answer accompanied by its sources: document name, knowledge level it comes from, fragment used. The conversation settles into the client's long-term memory, isolated per user.
The infrastructure
What the Cognitive Layer contains.
UBcore lives at the Cognitive Layer, above the client's transactional systems and operational automations. The infrastructure that follows describes what the Cognitive Layer contains, from the inside.
Three architectural components. Each is independent of the others: it can be replaced, extended, redistributed without touching the rest of the system.
ACCESS AND ISOLATION
Access to the system is described in the operational flow above. At the infrastructural level, the isolation of the client's data is a property of the architecture, not an application-level filter: the Security section explores this property in depth.
Permissions are granular, by role and by operation. The system's programming surface is documented and inspectable.
THE MEMORY
The heart of the infrastructure. The client's memory lives in three forms: the document store (the uploaded documents, isolated per client), the semantic memory (indexing by meaning), the structured Knowledge Base (the hierarchical organization of knowledge).
The structured KB is a first-class architectural asset. The Memory Cosmos section describes how it works.
Each client operates on their own memory instance, physically separated from the others. It is not logical segregation, it is infrastructural segregation.
THE REASONING
The architecture is agnostic to the reasoning engine. The system can work with commercial frontier models for cases that demand the highest quality, and with local on-premise models for contexts where data cannot leave the corporate perimeter. The choice of model belongs to the client, not the system.
The model access credentials stay with the client. The system handles them in transit, encrypted, for the duration of the single request.
Changing model is a configuration change, not a code change. When a provider releases a more powerful version, the system's performance follows automatically.
Security
Filter first, not only after.
Most AI systems check the answers. UBcore checks what enters the client's memory. Security is not a filter on the output, it is a property of the input.
PRE-INGESTION SENSE FILTER
Every document that enters the system is read by a dedicated model before being indexed. The model assigns a risk score on a graded scale. Documents destined for the strategic truths, where the company constitution and the non-negotiable criteria live, must pass a higher critical threshold. If they do not, they are blocked. The constitution does not get contaminated. Ambiguous content destined for the operational levels is flagged for human review. The system does not decide on its own what is admissible at the threshold, it asks.
STRUCTURAL ISOLATION, NOT APPLICATION-LEVEL
Three forms of separation operate at the same time: separation of base data, separation of semantic memory, separation of archived documents. Each client lives in their own space, not shared, not logical. The segregation is not a rule the software applies, it is a property of the architecture. No client can see another client's space, not even through a configuration error.
TRACEABILITY BY CONSTRUCTION
The traceability of responses is described in the operational flow. As an architectural property: every system action leaves a trace. The system is designed to make every decision verifiable, in line with the principles of transparency and human oversight that the European regulatory framework requires.
Memory Cosmos
Knowledge, organized by priority.
The heart of UBcore is memory. Memory Cosmos is the way this memory is organized, because a memory that is not organized is worth little. A company's knowledge has a natural hierarchy, and Memory Cosmos recognizes it instead of flattening it.
The Three Levels of Priority
The three levels (strategic truths, operational documentation, project context) were introduced in the operational flow. Memory Cosmos implements them through a specific property: the pre-ingestion sense filter decides what enters which level, based on who deposits it and on a risk score assigned by a dedicated model.
When the system has to answer a question, strategic truths enter first, always. Operational documentation integrates. Project context anchors to the present moment. The result reflects what the company has consolidated over time, not what the model statistically remembers.
The Constitution, the Privileged Level
The strategic truths deserve separate consideration. The company constitution is that piece of knowledge that expresses who the company is, what it decides a priori, what it is not willing to negotiate. It is deposited by the client, written by the client, maintained by the client.
UBcore loads it into every response. It is not a reference the system can consult if needed, it is a level that enters upstream of every elaboration. A strategic decision that contradicts the constitution is flagged. A new position that forms coherently with the constitution consolidates with an explicit reference.
For the client it means having a tool that does not forget who it is, even when the pressures of the moment would push in a different direction.
The Sense Filter Above the Fragmented Landscape
No real company keeps its knowledge in a single place. Documents in collaboration environments, archives in internal document systems, notes in different knowledge management tools, attachments in email, reports coming out of management systems: a company's document base is fragmented by nature.
UBcore does not ask the client to unify upstream. It reads the sources where they are. On top of them, it applies its own internal hierarchy as a common grammar of reading. Whoever indexes a single source knows that source. UBcore works on top of the client's landscape as a whole, reading it in the language of the company.
The memory that results is not a copy of the documents. It is a network of semantic relationships, which makes it possible to follow how a position formed, how it was updated, where it was superseded.
Where Knowledge Comes From
UBcore touches by nature four types of sources.
Type A are the client's document environments, where textual knowledge already exists. UBcore reads them directly, with the protocols the market has consolidated for this kind of access.
Type B are the client's transactional systems, where the operational data of the business lives. UBcore does not connect directly, it receives the flows the reasoning needs through a defined receiving point, where client and system partner establish what passes and in what form.
Type C are the client's operational automations, where its agents produce structured output. UBcore receives this output from the same receiving point, integrating it into its own memory with the same filtering discipline.
Type D are external vertical domain sources, sector databases, regulatory sources, specialized repositories. UBcore receives them in normalized form, through the same mechanism.
Type A is the only case of direct reading. Types B, C and D share the same pattern: a defined surface where client and system partner build the connection, with clear technical and contractual rules. UBcore does not build industrial connectors toward the client's transactional systems. It defines the surface on which those systems deliver their flows.
Decision Quartet
Four ways of never trusting a single model.
The central problem of an AI system that has to support decisions that matter is trust.
Not the romantic trust in the model, the kind marketing promises when it says «zero hallucinations». Operational trust: the decision-maker must be able to contest the answer, see where it comes from, understand where it can be wrong, compare it with another. To be used in real decisions, an AI answer must be put under tension.
UBcore operates through four independent patterns of putting reasoning under tension. Four ways of never trusting a single model.
Parallel Fact Checking puts the facts under tension. When an answer requires adherence to external reality, the system compares what the model asserts with what web search returns at the same moment. The model's memory against reality right now.
Dialectic puts thought under tension. On cognitive choices that require critical evaluation, the system builds the argued thesis, sets it in structured opposition with its own antithesis, returns to the decision-maker a synthesis with explicit conditions. The model against itself.
Parallel Reasoning puts the voice under tension. On decisions where the slant of a single model can carry an invisible bias, the system queries two or three different reasoning models in parallel, and an impartial judge synthesizes convergences and divergences. The provider against other providers.
Parallel Deep Research puts the picture under tension. On decisions of weight that require articulated evidence gathering, the system combines multiple reasoning engines with iterative multi-step web search, and a judge integrates the overall picture. The evidence against its own partiality.
The four patterns are activated in different ways. Parallel Fact Checking operates as a first anti-hallucination defense, accessible to every configuration. Dialectic activates automatically when the system recognizes a cognitive decision of weight, with the possibility of manual override. Parallel Reasoning and Parallel Deep Research are manual, activated by the decision-maker when the stakes require it.
The system is agnostic to the reasoning engine and supports different modes of provider access, from single-provider configurations for compact realities to multi-provider configurations with maximum sovereignty for enterprise realities.
An Orchestrating Agent, Under Construction
Above the four patterns of the Decision Quartet, UBcore is building an orchestrating agent: it uses the four patterns as tools of a longer reasoning, activating them according to the nature of the decision to be made.
It is a work in progress declared openly, not a promise. It is not yet demonstrable. We tell it as it takes shape, on UBcore's channels.
What the Decision Quartet does today is already enough to sustain real conversations with real decision-makers. The orchestrator is added when it is ready.
Foundational Layer
First the grammar, then the reasoning.
A company's knowledge exists, but it lives scattered: in document systems, in emails, in individual spreadsheets, in shared drives, in archives no one opens anymore, in conversations no one transcribes. Asking an AI system to reason on this mass without structuring is like asking a consultant to walk into a messy warehouse and pull out a strategy in half an hour. It can be done. It is worth little.
The Foundational Layer is the path UBcore builds with the client before switching on the Cognitive Layer:
- Initial discovery: the system listens to the client, gathers what already exists, maps what is there and what is missing.
- Strategic constitution: the client's leadership formalizes its non-negotiable principles, the criteria of choice, the consolidated positions. It becomes the level of strategic truths.
- Document selection: the foundational documents are identified, the ones that deserve to enter the strategic truths of memory. Few, chosen, passed through the sense filter.
- Operational expansion: the documentation of daily work enters the system under quality control. No noise, no duplicates that contradict each other without order.
- Project context: the work streams in progress populate the living level of memory, where decisions are still taking shape.
The Foundational Layer is not external consulting. It is a component of the product. Without the Foundational Layer, the Cognitive Layer works on noise; with it, it works on grammar.
For the client it is also a training path. Working with UBcore on structuring its own knowledge, the company rediscovers what it has, what it has lost, what it decides a priori. Often it is the first time it looks at it together.
The vision
The brain above the arms.
Today's market invests in the arms: agents that generate leads, workflows that answer emails, bots that produce reports. It is operational automation, immediate productivity.
UBcore does not automate. It governs. Automation without governance scales chaos. Governance without automation stays an intention. UBcore holds the two together.
The next step
Now you know the infrastructure.
If what you have read corresponds to a problem you are facing, the next step is a conversation.
LET'S TALK →