When a company loses the why, it loses its memory

The conclusion survived, but its conditions did not

Imagine a small business appointing a new sales manager. The files contain a pricing policy, meeting notes, a customer list and exchanges with management. A substantial discount granted three years earlier looks like a useful precedent for a difficult negotiation.

This is a fictional scenario. The discount compensated for a supply disruption that prevented the business from meeting a commitment. It was authorised for that customer and later superseded by a different rule. The amount survived; its scope was less well preserved. Copying it would turn an exception into sales policy.

The problem exists independently of AI. It starts when someone mistakes a recovered conclusion for a decision that still applies. Keeping information does not necessarily preserve the conditions under which it may be used.

Connecting what makes understanding possible

Organisational memory has a research history: Stein and Zwass proposed an information systems model for it in 1995 [1]. Here I propose a practical guide, rather than a scientifically validated model: record → context → status → knowledge → possible action.

A record may be an email, meeting, photograph or documented decision. Context identifies the people, time, purpose, constraints and alternatives considered. Status tells us what we actually have: a documented fact, decision, testimony, hypothesis, interpretation, disputed or superseded information, or an unresolved question.

Knowledge is what we can learn from these connections. Acting on it requires another judgement: does that lesson apply here, today, to this person? A retained decision does not automatically confer permission. Preferences change; procedures may apply only under particular conditions.

In our business, contextual memory should be able to say: ‘This discount was approved on this date to compensate for this problem. It applied to this customer and was later superseded by this rule.’ Each element should link to evidence. If the reason is missing, the gap must be acknowledged, not filled with a plausible explanation.

Transfer knowledge before the expert leaves

Tacit knowledge also calls for practice. A 2026 study at Fagor Automation describes learning through supported tasks; the intervention was only partly successful [2]. It does not establish that comprehensive paperwork alone can secure a succession.

My practical proposal is to let the successor work through a real decision while the expert can still observe and explain. Hesitation reveals unwritten criteria, abandoned attempts and people worth consulting. Recording those explanations preserves the reasoning, while attributing recollections gathered after the event.

Recording who produced what provides another foundation. W3C PROV offers a provenance framework connecting entities, activities and agents [3]. It does not certify a statement’s truth or its current commercial validity. Those judgements still need documentation and governance.

AI expands the reach of old decisions

AI makes large volumes of records searchable. Yet retrieval based only on similarity could present our discount as valid practice: it resembles the request, while its former scope goes unrecognised. A hypothesis can become a fact, a summary a source, and a particular experience a general recommendation.

The Memora benchmark, published in Findings ACL 2026, evaluates the reuse of invalidated memories in synthetic conversations [4]. MCMA experiments with abstraction levels in ALFWorld, ScienceWorld and BabyAI [5]. These experimental results do not establish the reliability of a real sales agent.

My interpretation is that useful memory must preserve both the status of what it contains and the level at which an experience can be reused. An accurate recollection can become harmful when generalised without justification. Useful forgetting includes withdrawing a former rule’s present authority, without turning that distinction into a requirement to retain everything.

The more an AI can act, the more it must be able to justify its evidence. Here it should distinguish historical precedent from the current rule and leave approval of the discount to an authorised person.

Cognitive continuity as an ambition

This is what I mean by ‘cognitive continuity’: a proposed term for carrying reasoning across changes of people and tools, rather than a universally established scientific category. Memory becomes an infrastructure for continuity when it also explains what has changed.

That ambition informs the principle pursued by Mnemory Knowledge: preserving relationships between records, context, status, time and knowledge. The purpose extends beyond another document repository. It concerns people, their successors, assistants and future AI agents, including after organisational change. This is a design direction to test, not a claim that all these capabilities are already guaranteed.

Put it into practice

Choose an important decision made more than a year ago. Try to answer five questions:

1. What decision was actually made?
2. Why?
3. What alternatives were considered?
4. Does the decision still apply today?
5. Could someone joining tomorrow understand all this without asking those who were there?

If the final answer is no, the business probably has the information, but not yet all the memory its continuity requires. This proposed test is a working exercise, not a scientific measure.

What remains usable

Storage preserves material. Continuity requires its meaning to remain accessible and its use open to scrutiny.

A company’s memory is measured not by what it keeps, but by what someone else can still understand and use when those who knew are no longer there.

Philippe Contal