One company memory can still produce five versions of the truth

Putting everything in one place does not create one reliable truth. Agents still need the right permissions, the current version and a clear record of what changed, or shared memory becomes shared confusion.

Here's a brief overview.

0:00

/93.019138

1×

Centralising company knowledge feels like the obvious answer to fragmented AI.

Connect the documents, index the messages, bring the policies together and give every agent access to the same memory.

The business now has one repository. It may still have five different truths.

One agent retrieves an old policy. Another sees the current document but misses the exception approved last week. A third has access to a broader set of records and reaches a different conclusion. A fourth relies on a summary that no longer reflects the source. A fifth cannot explain which version shaped its answer.

The problem is not simply where information is stored. It is how truth is identified, governed and carried into work.

One location is not one source of truth

A shared drive can hold several versions of the same plan. A knowledge base can preserve a policy beside the message that replaced it. A project system can show the latest status while a document still contains the previous decision.

People manage these contradictions through experience. They know which folder is current, who changed direction and which exception applies to this customer. An AI agent does not inherit that judgement merely because it can search the same material.

Access to more information can make the problem harder. The agent has a larger set of plausible sources, but no reliable way to know which one has authority.

Company memory therefore needs more than storage. It needs a clear relationship between the source, the decision, the approval and the work that follows.

Currency is part of meaning

Business information changes over time. A price, approval boundary, campaign direction or hiring plan can be correct on Monday and wrong on Friday.

If an agent retrieves the older statement without recognising that it has been superseded, the answer may be coherent and still be operationally unsafe. The more capable the model, the faster that stale context can travel into other systems and decisions.

Every important record should make its status understandable. What is current? What was replaced? When did the change apply? Who approved it? Does the new decision affect all work, or only one market, customer or period?

Without that context, the repository contains history but cannot reliably guide action.

Shared memory does not mean shared access to everything

Truth is also shaped by permission.

Two agents may be working from the same company memory but should not necessarily see the same information. A customer-support agent needs the approved policy and relevant customer context. It does not need unrestricted access to every finance, legal or employee record.

The goal is not universal visibility. It is relevant, authorised context.

Role-based access reduces exposure, but it creates another requirement: the business must understand why different agents may reach different conclusions. If one answer depends on a restricted source or a specific exception, that boundary should be visible to the accountable person.

A trustworthy system can distinguish “I found nothing” from “I am not authorised to use that source.” Those are different conditions and should not produce the same confidence.

A decision needs a receipt

When an AI-assisted action matters, the business should be able to trace the basis for it.

That does not mean exposing hidden model reasoning. It means preserving the operational evidence that a person would need to review the decision:

• the sources used
• the version or effective period
• the standing rule or approval boundary applied
• the exception, if any
• the person or role accountable for the outcome
• the change made in the business system

This record makes disagreement diagnosable. If two agents produce different answers, the team can compare their evidence and permissions instead of debating which response sounds more confident.

It also allows corrections to improve future work. A resolved conflict can become a governed decision rather than another note buried in a chat.

The Walter point of view

Company memory should function as an operating asset, not a passive archive.

The requirement is to connect current company context with clear authority and accountable action. Walter is designed around that principle: helping people and agents work from governed context, maintain continuity across hand-offs and keep the relevant decision history close to the work.

The value is not that every agent repeats an identical sentence. Different roles may need different views. The value is that those views are drawn from controlled sources, respect permission boundaries and can be reconciled when they diverge.

One company memory becomes useful when the business can explain not only what an agent knew, but why that knowledge was current and appropriate for the action.

Test the integrity of one company truth

Choose a decision that affects several teams or agents. Ask each workflow to identify the current rule, its source, its effective period, any exception and the accountable owner.

If the answers differ, do not begin by tuning the prompts. Find the break in the truth chain. The source may be stale, the authority unclear, the permission boundary incomplete or the correction missing from the shared record.

Centralisation is a useful first step. Integrity is what makes the memory dependable.

The goal is not one enormous store of company information. It is one governed way for the organisation to know what is true now, who may use it and how the next action should remain accountable.