Context Fabric

The problem it solves, in the provider's words

Context Fabric turns what employees tell AI assistants into governed organisational context: every stated fact is proposed, then accepted or rejected by someone accountable, carries assurance and provenance, and is erased together with its source.

The problem

Employees describe their organisation to AI assistants all day: which data may leave the EU, which identity provider is approved, what the security team requires. The assistant remembers. Left untreated, that memory quietly becomes the organisation's working "truth" — nobody accepted it, nobody can say who said it, a later contradiction overwrites it without trace, and when a source must be deleted, the derived statements live on.

The business rules

• A stated fact is a proposal until someone accountable accepts it; it can be rejected.
• Acceptance has levels of assurance: inferred, user-declared and source-backed are recorded as different things. A fourth level, "verified", exists in the vocabulary and is deliberately unreachable today.
• Facts change: a newer statement supersedes the older one, which stays on record; statements can be withdrawn. Nothing is silently overwritten.
• Everything traces to who said it and where, including the conversation session and model run that produced it.
• Erasing a source erases what was derived from it — text, chunks, embeddings and quotations copied into stored decision records.
• Business authority is never model-instruction authority: a corporate policy can be evidence; its text is never an instruction to the model.

The pieces, and how they fit together

• Context Profiler (live) — turns what is said into proposed assertions, each with attribution and conversational provenance. Nothing it produces is organisational truth by itself.
• Accountable acceptance (live) — a person on your side promotes each proposal to accepted or rejected: one atomic operation, audited in the same transaction.
• Governed context store (live) — assertions with assurance levels, sources with sensitivity class and permitted operations, and immutable digest-stamped snapshots at material events. Supersede and withdraw keep full history.
• Context Envelope (live) — every AI call receives a bounded set of eligible context, never open access to the store. The full multi-step Context Composer (resolver) is roadmap.
• Reference Knowledge (live, companion subsystem) — external reference material with profiled sources and provenance, feeding grounded advice into the envelope alongside organisational context. The two are never confused: reference knowledge informs, context declares.
• Protection boundary (live) — registered client and third-party identities are replaced with stable stand-ins before derived context is stored.
• Erasure and data-subject rights (live) — withdraw (reversible) and erase (destroys content and every derivative) are separate operations; snapshots are redacted in place; export covers the context tables.
• Context-use evidence (live) — per model turn: what context was considered, supplied, dropped (with the reason) and relied upon.

The mental-model diagram in Media shows the whole flow, with live pieces and roadmap marked.

Where it applies, beyond a chat assistant

• Agent fleets acting on company facts — each action traceable to an accepted, current assertion.
• Compliance and procurement copilots — constraints attributable and withdrawable.
• RAG and knowledge systems that must honour "forget this source" through every derivative.
• Vendor onboarding and due-diligence rooms — who claimed what, on which basis, superseded by what.
• Any regulated workload where an answer must show the context it rested on.

The three limits that matter most

1. Proven inside AI Exchange only — no second consumer exists yet; portability is roadmap.
2. "Verified" is never claimed — all assurance today is inferred, user-declared or source-backed; verification independent of the builder does not exist yet.
3. Pre-production evidence with synthetic data — production operation is not claimed.

(The full list of nine declared limitations remains below and in the declarations.)

Where it matters most

Regulated industries (financial services, health, public sector), procurement and vendor selection, and any organisation subject to data-subject rights, where "forget this source" must mean something checkable.

The provider's value hypothesis

The problem

AI assistants can turn informal organisational statements into durable working truth without accountability, provenance, currency or erasure.

Who benefits

Enterprises whose AI assistants accumulate institutional knowledge, especially regulated industries and procurement/vendor-selection workflows.

What it provides

A governed context layer that records proposals, acceptance, assurance, supersession, provenance and erasure of derived data.

It could reduce the risk that chat memory becomes organisational fact by default, while making answers auditable for the context they used and allowing accepted context to be reused across sessions.

Value drivers

• avoids: uncontrolled chat memory becoming organisational truth
• reduces: re-asking settled constraints across sessions
• avoids: untraceable contradiction overwrites
• reduces: loss of auditability over what an answer rested on

Hypothesis declared by the provider — not independently verified; figures marked illustrative are examples.

What it does

Capabilities, as classified on AI Exchange

• AI Context & Memory Governance
• Data Governance
• AI Decision Traceability & Audit

Where it applies

Industry

• Cross Industry

Business function

• Information Technology
• Risk & Compliance

Business process

• Knowledge Support

Outcome

• Improve Compliance Readiness
• Reduce Risk Exposure

Will it fit your environment?

What it needs from you

• Human Final Approvalrequired

What it depends on

• Provider-neutral model gateway (Azure OpenAI deployments today)servicerequired

Declared by the provider. Fit compares these against your own constraints.

Need it adapted to your environment?

Contact the expert

What it will not do

• Proven inside AI Exchange only; portability is not claimed.high
• No certification or regulatory attestation is claimed.low
• Protection is recall-limited and covers identities the classifier names.medium
• Supplied originals remain in scope as supplied and are not transformed by the protection boundary.medium
• The full multi-step resolver is roadmap only.medium
• Does not verify organisational facts (the verified assurance level is deliberately unreachable); does not arbitrate conflicting statements; does not set retention periods

Not for

• A standalone user interface for the protection boundary.
• Not a DLP or content-detection tool; protects registered identities only

What adoption involves

Indicative price

From €20,000

Defined-scope pilot (one client, one evaluation environment, one principal flow, bounded integration; approx. 4-8 weeks subject to requirements). Indicative only; see "What you receive". Larger deployments and support are scoped separately.

Implementation time

Custom / depends on scope

Depends on which assistant surfaces consume governed context and on the customer-side acceptance workflow

How it runs

Implementation — a consequence of the rules, not the core

Current binding: PostgreSQL-backed context schema inside the AI Exchange platform, integrated with its Discovery and Builder assistants, hosted on Azure. Model calls go through a provider-neutral model gateway; protection-key custody is managed configuration separate from database access.

Portability (two-tier):

• Golden v0 — this publication: proven inside AI Exchange, for AI Exchange's own workloads.

• Portable tier — planned next: extraction of the governed-context layer for consumers outside AI Exchange is the declared next step on the provider's roadmap after the current platform freeze. The generic context-use layer is already kept free of marketplace semantics by design; the remaining work is extraction and a second consumer.

Deployed evidence is from AI Exchange's pre-production environment (which serves its public domain) with synthetic data; production is not claimed.

Adaptations

Not part of the terms above. Whether it can be adapted, and what that changes in price and timeline, is the provider's answer to give.

Contact the expert

Provider, evidence and what is still unknown

Provider

AI Exchange · Operates this platform · Verified organization

What does verified mean?

Evidence and assurance

Assessment from golden-v0 (Core v1 slice) · material unchanged

Evidence statements: 5 provider-declared.

Automated checks were run on the submitted material.

Detailed assurance information is available to signed-in buyers.

No certification or security approval is implied.