Cortium
The next bottleneck in AI is not intelligence. It is judgement.
Models can already search, reason, write, and act. The harder problem is helping them understand how a company decides what matters before they do anything consequential.
A note on the future we are building towards, and Cortium’s place in it.
Institutional memory → grounded actionWhy we are building Cortium
Everyone is working on how models think. We are interested in how organisations decide.
A company can buy better models, connect more tools, and automate more work. It can still make the same judgement call five different ways because the reasoning behind previous decisions was never kept.
The important knowledge is often not the final answer. It is why an exception was allowed, which precedent mattered, what risk changed the decision, and what happened afterwards.
Today, that knowledge is scattered across Slack threads, tickets, pull requests, meetings, and the memories of experienced people. Search can find fragments of it. A model can summarise those fragments. Neither gives the company a durable way to decide.
We believe companies need a decision layer: shared memory for the judgement that sits between knowledge and action.
A working definition
What is a decision layer?
A decision layer is the system that turns company context into a recommendation that can be explained, reviewed, and acted on.
It does not replace data, search, or workflow systems. It gives them a consistent standard of judgement to work from.
Five convictions
The bets behind Cortium.
- 01
Intelligence gets cheaper. Context gets more valuable.
Models will keep improving. The lasting advantage will come from knowing how a particular company interprets policy, weighs trade-offs, and handles exceptions.
- 02
Agents will create more decisions, not fewer.
When software can act, every action carries a judgement call. Faster execution only makes the quality and consistency of those decisions more important.
- 03
The most useful company knowledge is rarely written down.
Documents explain the rule. Real judgement appears when the rule meets an unusual customer, a production incident, or a decision nobody has seen before.
- 04
Memory has to live outside the model.
A prompt is temporary. A company needs persistent memory that can be inspected, corrected, governed, and used by whichever model comes next.
- 05
Autonomy should be earned.
A system should recommend first, expose uncertainty, learn from review, and automate only the decisions it has shown it can handle reliably.
The judgement stack
Four systems. Four different questions.
These layers work together, but they are not interchangeable. A company can have excellent data, search, and automation while still lacking a repeatable way to make judgement calls.
Data
What happened?
Transactions, events, messages, and state.
Knowledge
What do we know?
Documents, definitions, policies, and retrieved facts.
Workflow
What happens next?
Tasks, approvals, routing, and execution.
Decision
What should we do, and why?
Context, precedent, uncertainty, review, and outcome.
Search can find the policy. Workflow can route the approval. The decision layer explains whether the exception should be made.
Knowledge → judgement → actionThe missing memory
Documents preserve what a company knows. Decisions preserve how it thinks.
A policy says refunds close after thirty days. It does not explain why the company made an exception for a long-term customer after a service outage.
The exception contains the useful judgement. Context, policy, precedent, review, and outcome belong together as one record.
Enough records create something documents cannot: an evolving model of how the organisation handles uncertainty.
A lesson from agent research
The Context Gathering Decision ProcessA larger context window is not the same as reliable memory.
Long-running agents can lose their objective, repeat the same search, or stop after finding a plausible answer. The paper frames this as a context-gathering decision problem.
Its useful lesson is architectural: keep important state explicitly, track unresolved questions, and use controls outside the model to decide when enough evidence exists.
- 01 ObserveGather evidence from the environment
- 02 UpdateKeep a compact, explicit belief state
- 03 ResolveTrack what is known and what is still missing
- 04 DecideAct, abstain, or gather more context
Cortium applies the same principle at company scale: important judgement should be persistent infrastructure, not something a model has to reconstruct from scratch every time.
What this means for Cortium
Cortium should become the memory people and agents consult before they act.
Capture decisions, not noise.
Find the moments where judgement changed what happened, then preserve the context around them.
Ground every recommendation.
Cite the policies, people, and comparable decisions that support a proposed action.
Make uncertainty visible.
Confidence should determine whether the system answers, abstains, or routes the decision to a person.
Learn from what happens next.
Reviews, corrections, and outcomes should make future decisions more consistent and more useful.
What we do not want to build
Not another assistant that sounds informed. Not surveillance disguised as knowledge management. Not an autopilot that hides uncertainty behind a confident sentence.
The purpose is not to remove human judgement. It is to stop good judgement from disappearing, repeating, and becoming a bottleneck. People should remain able to inspect the evidence, change the decision, and decide where automation belongs.
Where this goes
The model is replaceable. The memory a company builds is not.
We are building Cortium to become the shared decision memory beneath a company’s people and agents. A place where judgement compounds, autonomy is earned, and every consequential action can be traced back to the reasoning that made it possible.
Build this with usReferences
A few of the ideas and standards informing this direction.
- 01 Read paper
The Context Gathering Decision Process
Why reliable agents need explicit state and programmatic controls beyond a growing prompt.
- 02 Read framework
Artificial Intelligence Risk Management Framework
NIST’s framework for governance, measurement, transparency, and human oversight.
- 03 Read principles
OECD principles on transparency and accountability
A practical standard for traceable systems and decisions that people can understand and challenge.
- 04 Read overview
The European Union AI Act
The official overview of risk-based obligations, including traceability, monitoring, and human oversight.