Operating model
A context-window lens for organisational memory.
The same shape, scaled up
A context-window lens for organisational memory
A company does not literally have a context window. People are not tokens, and organisational memory has no single numeric limit. Still, the analogy is useful. A language model can act only on the instructions, evidence and state available in its current window. A team can decide only from what its people can find, trust and bring into the room at the right moment.
Most company knowledge is outside that active working set. A roadmap decision may live in a former employee's memory, an old message thread and a deck whose status is unclear. The information exists, yet the person making today's choice cannot use it confidently. That resembles a context problem: not an absence of information, but a failure to make the relevant, current and authoritative part available when the task needs it.
What should remain resident in organisational memory
Use the context-window analogy for selection, while preserving the human and political limits the analogy cannot capture.
- Current goals and constraints What the organisation is trying to achieve now.
- Consequential decisions Choice, rationale, owner and scope.
- Primary evidence Sources that can support or challenge the decision.
- Review triggers Conditions that reopen the memory.
- Protected or tacit knowledge May require restricted access or human conversation, not broad retrieval.
Useful memory is selected and governed, not accumulated without limit.
The failure modes line up
Sometimes the friction is a context problem in disguise
Consider a vendor renewal. The procurement record says “renew for three years,” but not why. Finance remembers a discount. Security remembers an unresolved exception. Operations knows a migration was postponed because a customer deadline took priority. Each person holds a legitimate fragment, and no single document shows the conditions that made the earlier decision sensible.
The team can repeat the old conclusion, relitigate the choice from scratch, or reconstruct the decision. In this case, reconstruction is better: retrieve the signed terms, approved risk exception, cost model and customer dependency; identify an owner for each; then state what has changed. The result might be “renew for one year while the migration completes,” with evidence and a review trigger. The value is not perfect memory. It is a decision whose reasoning can be inspected and revised.
The analogy helps classify other friction too. Onboarding assembles the minimum working set a new colleague needs. Meetings synchronise context when a document cannot carry the disagreement. Silos make relevant evidence inaccessible at the point of decision. Repeated analysis can mean an earlier answer lacked provenance, a clear record of where it came from and who approved it, was impossible to find or no longer fits the facts.
These are not always context failures. A meeting may be needed for trust, negotiation or accountability. Repeating analysis can be responsible when the world has changed. And some knowledge is tacit: learned through practice, relationships and judgement rather than captured fully in text. The analogy is a diagnostic, not a reason to treat people as storage devices.
Organisational forgetting has several causes
Not every memory failure is a retrieval problem, and not every answer is more storage.
Memory-system failures
- Missing provenance A claim cannot be traced or dated.
- Stale authority An old decision still appears current.
- Permission leakage Sensitive knowledge reaches the wrong audience.
diagnose the failure class
Organisational failures
- No decision owner Nobody can resolve disagreement.
- Tacit knowledge excluded The system stores documents but misses practice and context.
- Power hidden by consensus A neat summary erases who was affected or overruled.
Fix governance and relationships where retrieval cannot solve the problem.
You already know the fix
Harness engineering at company scale
The model discipline offers useful practices, not a complete corporate blueprint. Harness engineering asks what evidence enters a task, which tools are available, what state persists, how a result is checked and where authority stops. At company scale, those questions become source ownership, access rules, decision records, review dates and accountable approval.
For an important decision, preserve five things: the conclusion, the evidence, the owner, the conditions under which it holds and the trigger for review. Make that record searchable, but do not flatten every source into one confident summary. Conflicts and uncertainty are part of the context too. This is the governed version of a company brain: memory that can be challenged, corrected and forgotten when it no longer has a lawful or useful reason to remain.
Make the decision inspectable. Preserve its evidence, owner, conditions and review trigger.
What it's for
A managed window for the whole organisation
This is one purpose of central intelligence: connect current work to governed organisational memory. “Central” should describe a shared way to find authority and provenance, not one database or one team deciding what everyone must believe. Legal, finance, operations and sales may keep different systems and permissions while using common rules for evidence and review.
The practical test is simple. When a decision repeats, can the next person find the earlier reasoning, see whether its conditions still hold and identify who may revise it? If not, the organisation is paying to reconstruct its own past. Fix the retrieval and ownership problem before adding another summary tool.
Where the analogy ends
Manage information without reducing people to context
A model's context is supplied to it. People interpret, disagree, withhold, forget and create new understanding. Power matters: deciding which record is “authoritative” can silence the colleague closest to the work. Privacy matters too. Not every conversation should become permanent organisational memory.
The student lesson is to use the analogy until it stops helping. It reveals why evidence must be available at the moment of decision. The professional responsibility is to add consent, ownership, challenge and deletion, because an organisation is not a larger language model. Managed well, its memory does more than save time. It lets people spend less attention reconstructing the past and more judgement deciding what should change next.