A cognitive infrastructure is not delivered: it is operated. And operating is not
“keeping the server on”: it is managing the lifecycle of the
knowledge, because the world it describes does not stop moving. An ICE
that is not updated is not an infrastructure: it is a map of the past that
gives false confidence.
Introduction
Yesterday we closed governance: the agent acts and can be audited. But there is
a problem that governance on its own does not solve, and it is the one that gives the
title to this article: knowledge changes.
The rules are revised. Suppliers are renewed or disappear.
Contracts are renewed. Org charts change. Prices are updated.
Regulations evolve. And meanwhile, the cognitive infrastructure that
we have described for twenty days contains a photograph of that world.
The question, then, is not “do we build the ICE?” (that is already solved). It is:
“how does it stay alive when the world it describes does not stop moving?”
And the answer is a cycle, not a state: the knowledge lifecycle, the
lifecycle of knowledge.
The problem
A logistics operator has had its cognitive infrastructure over
the supplier and contract network for a year. It worked. Until it did not:
- “The agent said that supplier X is still active, but X entered bankruptcy
last month. The ICE has it as an in-force supplier.” - “We renewed the contract with Y in March, but the agent keeps applying the
prices of the old contract.” - “We closed the relationship with Z last year, but it keeps appearing in the
supply chains. Nobody eliminated it.” - “The price of component W changed in the ERP, but the
knowledge layer still has the price from six months ago.”
Four failures, and the four have the same origin: the ICE does not have a
lifecycle. There is no process that incorporates the changes, validates them,
versions them, updates them and, when the time comes, expires and eliminates them. The result is
an infrastructure that desynchronizes from the world, and that — worst of all — does it
in silence: the agent keeps answering with the confidence of always,
over a past that no longer is.
The problem is not that knowledge changes: it is inevitable. The problem is
that there is no mechanism for the change to reach the ICE in a
controlled, validated and auditable way.
The concept
The knowledge lifecycle is the set of states a
piece of knowledge passes through from birth to death, and the processes that
govern each transition. In an ICE, the cycle has nine states:
- Incorporation (ingest). The knowledge enters the ICE: from a
source system (ERP, CRM), from a document, from an event of the
operation, or from an agent. “The new contract with Y is incorporated.” - Validation. It is checked that the knowledge is correct, coherent and
from an authorized source, before publishing it. “Is the contract with Y
signed? Are the prices the agreed ones? Does it not contradict another rule?” Without
validation, garbage is incorporated. - Publication. The validated knowledge becomes available to the
agents. “The new contract with Y is now the in-force one.” - Use. The agents consult and use the knowledge. It is the
normal operational state. - Update. The knowledge changes (new price, new term) and
is updated, keeping the traceability of the previous. “The price of W
goes from 12 to 14 €.” - Versioning. Each change is versioned: you can see which version
was in force at each moment. “The contract with Y: v1 (2025), v2 (2026).”
Versioning is what allows answering “what rule applied on April 3?”
(Day 18, temporal validity). - Correction. An error is detected and corrected, with traceability of the
correction. “X is not in bankruptcy; the state was corrected.” The correction
does not erase the error: it leaves it registered. - Expiration. The knowledge stops being in force (a contract ends,
a supplier is renewed with other terms). “The old contract of Y
expires on 31/3.” - Elimination + audit. The expired knowledge is archived or eliminated
according to retention, but it leaves a trace that it existed and why it was
eliminated. “The relationship with Z was eliminated on 31/12/2025 due to the
end of contract; it remains in archive.”
And over the cycle, a loop that keeps it alive: the feedback of the
operation.
KNOWLEDGE → (is used) → FEEDBACK → (is updated) → UPDATE
→ VALIDATION → NEW KNOWLEDGE
The feedback loop is what turns the ICE from a photograph into an
organism: the operation generates signals (a supplier fails, a price
changes, a rule is applied badly, an agent detects a contradiction), those
signals become updates, the updates are validated, and
the knowledge renews itself. Without the feedback loop, the ICE is static;
with it, it learns from the operation.
The key difference: the lifecycle is not “updating the
database”. It is a system with owners (ownership, Day 18), with validation,
with versioning and with audit, that manages knowledge as a
living asset. And it is the mechanism by which the ICE does not desynchronize from the world.
Architecture
The lifecycle is another transversal property of the ICE, like
governance (yesterday):
KNOWLEDGE LIFECYCLE
(incorporation → validation → publication → use →
update → versioning → correction →
expiration → elimination/audit)
+ operation FEEDBACK loop
┌──────────────────────────────────────────────┐
│ AGENTS · REASONING · CONTEXT · MEMORY │
│ KNOWLEDGE · ONTOLOGY · SEMANTICS · DATA │
└──────────────────────────────────────────────┘
The lifecycle governs the knowledge; governance (Day 18) governs
the decision. They are complementary: the lifecycle guarantees that the
knowledge is fresh and correct; governance guarantees that the
decisions that use that knowledge are auditable.
Use case
The logistics operator, with its ICE of suppliers and contracts, now with
lifecycle. We break down the supplier X failure:
- Problem. The ICE has supplier X as active, but X entered
bankruptcy. The agent keeps using X in the supply chains. - Feedback signal. The credit system (external source) emits a
notice: “X in creditor bankruptcy”. Or a risk agent detects it when
consulting X. - Incorporation. The state “X in bankruptcy” is incorporated into the ICE as
a state change of the supplier X. - Validation. It is validated: is the source reliable? (credit system,
authorized). Is it coherent? (there is no other record saying X is
active). Does it affect in-force contracts? (yes, two). - Update + versioning. The state of X goes from “active” to “in
bankruptcy” (v2). The change is versioned and traceable. - Repercussion (use). The agents that use X in supply chains
receive the signal: “X in bankruptcy → propose alternatives”. The graph of
relationships (Day 6) allows finding the alternative suppliers that
cover the same routes. - Correction (if applicable). If the bankruptcy is lifted, the state is
corrected back, with traceability. - Expiration + audit. If X definitively leaves the business, the
relationship expires and is archived, leaving a trace of the relationship and its end.
And the feedback loop: the bankruptcy notice (signal) → state change
(update) → validation → repercussion on the agents (use). The ICE did not
wait for someone to update it by hand: the loop kept it up to date.
Trade-offs
- Cost. The lifecycle is continuous infrastructure: ingest pipelines,
validation, versioning, and a team that manages the
updates. It is not built once: it is operated. - Complexity. Validating knowledge from heterogeneous sources (ERP,
documents, events, external sources) is complex: each source has its
own format, its own reliability and its own rate of change. - Concurrency. Several changes can arrive at the same time (a price changes
and a contract is renewed and a supplier enters bankruptcy). Managing
concurrency and the contradictions between changes is part of the design. - Validation vs. speed. Validating a lot is safe but slow; validating
little is fast but risky. In domains where a change is critical
(price, supplier, risk rule), validation has to be strict
even if it costs time. - Expiration and retention. What expires, how long is archived and what is
eliminated is a compliance decision (data retention), not only an
engineering one. An ICE that does not expire becomes a dumpster; one that
eliminates too much loses traceability. - Honest limitation: the lifecycle manages the knowledge that
reaches the ICE. If a change is not detected (a supplier enters
bankruptcy and no source emits it), the cycle does not incorporate it. The
coverage of the cycle depends on the quality of the sources and of the
feedback sensors.
Implication for the company
For the CDO, the knowledge lifecycle is what turns the ICE from a
project into a service: the asset is not delivered, it is operated. And it is an asset
with a dedicated team: without a person responsible for the cycle, the ICE desynchronizes
in six months.
For the AI team, the implication is one of design: the feedback loop has to
be designed from the start, not added afterwards. And measuring the freshness
of the knowledge (% of data updated in X days) is a production metric,
not an extra.
For operations, the message is the most tangible: the ICE stops giving
answers from the past. When a supplier changes, a price is updated or
a contract is renewed, the agent knows it. And that is what separates a
tool from a living capability.
For top management, the message is one of sustainability: the investment in the
ICE does not end with the deployment. It is a continuous investment, and its return
depends on the knowledge being kept up to date. An outdated ICE is not
an asset: it is a risk that gives false confidence.
Conclusion
The knowledge lifecycle is the set of states (incorporation, validation,
publication, use, update, versioning, correction, expiration,
elimination/audit) and the feedback loop that keep the ICE up to date
with the world. Without it, the cognitive infrastructure is a photograph that
becomes outdated in silence; with it, it is an organism that learns from the
operation.
We have covered the complete series: from data to knowledge, from
semantics to ontology, from entities to relationships, from the graph to the
memory, from context to reasoning, from the agent to the action, from
governance to the lifecycle. And the whole series has pointed to one thing:
building the layer that allows AI to understand an organization.
Tomorrow we close the series with the question that joins it all: what does this
mean for the future of the company?
Next article
Tomorrow: The Company After the Chatbot: Toward a New Cognitive Infrastructure. The closing article: not a summary, but a vision — the
evolution from ERP to data platforms to knowledge systems to cognitive
infrastructure to agentic enterprise — and the final question that the series
leaves open.
Sources
- Working definition of the ICE: see series, Day 16.
- The knowledge lifecycle as knowledge management with validation,
versioning and expiration is coherent with the content of Cirtra on
Knowledge Lifecycle Management (Agentic Knowledge Systems).
