Organisational memory: making knowledge outlive the chat and the analyst

Saving files is not the same as preserving intelligence

A folder of presentations, emails, prompts and newsletters is an archive; it may not be operational memory. To reuse a conclusion you need to reconstruct its provenance: which question it tried to answer, which sources were consulted, when, which criteria were applied, which alternative hypotheses were considered and which decision was made. Without that trail, an answer that is correct today can become an obsolete claim tomorrow.

We distinguish four objects that should be stored without mixing them:

1. Evidence: a primary document or record, its date, provenance, link and limitations.
2. Interpretation: a human or AI-assisted inference, its author, date and reasoning.
3. Product: the version delivered to an audience, with an explicit snapshot in time.
4. Decision and learning: the action taken, the conditions observed and the outcome when it can be measured.

You do not need to deploy a sophisticated repository to get started. A well-governed spreadsheet, a shared folder and naming conventions are enough for an initial memory. The right tool depends on confidentiality, size, permissions and availability.

Versioning is a promise of honesty

Living intelligence is not “silently corrected” as if no other conclusion had ever existed. A minimum history is kept: product ID, version, cut-off date, changes since the previous version, sources added or removed, interpretation modified, review owner and next update date.

Company A example:

• v1, 15/01: two preliminary research signals. Hypothesis: technical exploration; do not decide on an investment.
• v2, 15/03: one documented pilot test. Confidence changes that bounded operational validation exists; scale-up is still unproven.
• v3, 15/06: comparable operating data appears. The pilot assessment is reopened.

The value does not lie in the first answer having been “right”, but in being able to explain which evidence justified changing it.

What to do with AI and prompts

A prompt is an instruction, not a source. It is worth keeping the version of the methodology, the authorised inputs, the model or tool used when relevant, and the execution date. But an AI output does not automatically become primary evidence: it must link to verifiable records. Nor should confidential information be entered into external systems without the organisation’s authorisation and an assessment of their processing terms.

An additional risk is false memory: if an assistant summarises an earlier document with an error and that summary is reused as if it were a source, the error can accumulate. Every relevant product must allow you to go back to the original evidence.

Retrieval test · A conclusion that ages

Fictional file, Company A, P-NOR-004. In v1 the conclusion is “research and an announced pilot are observed; broad adoption is not demonstrated”. In v2 a pilot assessment adds a documented operating condition. The new information does not justify replacing “not demonstrated” with “adoption confirmed”. The level of detail must be modified, the limited conclusion kept and what is unknown explained.

Replacement exercise: another person, with no access to the author, must retrieve the original, reconstruct why v1 changed, tell the inference apart from the source and locate the access permissions. If they cannot, the memory is an archive that is hard to audit. The deletion or correction of personal data must be handled under the applicable retention policy, not as an exception to traceability.