An Institutional Memory That Can Be Wrong

Information used in coordination should retain enough structure to answer several questions: What is being asserted? Which source supports it? Who owns that source? When was it true? What cases does it cover? Is it a verified record, an inference, a proposal, or a disputed interpretation?

These distinctions matter because enterprise knowledge is not uniform. A signed contract and a draft policy can both be retrieved. A manager’s note and a system of record can both describe a project. Their presence in the same context window does not give them equal authority.

A future coordination architecture should preserve these differences through retrieval, summarization, and handoff. When an agent passes context to another agent, it should carry provenance and uncertainty along with the claim. Otherwise, each transformation can make a qualified statement appear more definitive.

Many organizational facts are time-dependent. A person’s role changes. An approval expires. A supplier’s capacity shifts. A customer commitment is amended. Even accurate information may become unsafe when used outside its validity period.

Memory systems should therefore represent time explicitly: when a claim became valid, when it was last confirmed, and when it should be reviewed or expire. Some claims need continuous refresh; others may remain useful as historical evidence. A system should distinguish “this was true at the time” from “this is currently true.”

That distinction supports both action and learning. A prior incident report may help explain why a decision was made without governing the next decision. A superseded policy remains part of institutional history but should not silently control current execution.

Correcting a source does not always correct the memory built from it. A summary may be cached. A downstream agent may have copied the claim. A decision may already have been made using it.

A reliable design needs a correction path that can locate derived uses, mark stale representations, and identify consequential actions that may need review. The goal is not to pretend that every error can be erased. It is to make the propagation of an error traceable and its correction actionable.

That requires records of which claims informed which decisions. The record should show the evidence considered, the authority exercised, and any uncertainty known at the time. This makes later review possible without rewriting the historical record.

Two authorized sources may disagree. Operations may report that an order shipped; the customer may dispute receipt. Finance may treat a milestone as achieved; delivery may say acceptance is still pending. A memory system that resolves the conflict by selecting whichever source is easiest to retrieve can convert a data problem into an unauthorized decision.

Instead, the system should preserve competing claims, explain their consequences, and route the conflict to the responsible authority. Once resolved, the decision should be stored with its scope and rationale. Future systems can use that resolution while retaining the fact that it applied under particular conditions.

This can make coordination slower in the moment. It can also prevent a confident but unsupported answer from triggering a costly action.

Today, organizations rely on source systems, document management, data catalogs, records policies, and human expertise to maintain institutional knowledge. Emerging agent systems increase the importance of connecting these controls to retrieval and action. The research does not establish that enterprises have solved this problem with a unified memory layer.

The practical design is likely to be distributed. Domains retain ownership of their records and policies. Shared coordination systems retrieve permitted claims with provenance, validity, and dispute status. Corrections flow to dependent uses where possible. Agents act only when the available memory satisfies the conditions set for the commitment.

A memory system earns trust not by appearing comprehensive, but by making its limits inspectable. It must show what it knows, where that knowledge came from, when it may no longer apply, and when the organization has not resolved what is true.

The central design test is consequential: when an agent relies on an outdated or mistaken claim, can the organization trace its use, correct the affected memory, and reconsider the decisions that followed?