Process Memory: Keeping Operational Experience Useful

What characterizes a process memory

The term builds on the concept of organizational memory. Walsh and Ungson describe its basic operation through the acquisition, retention, and retrieval of information from an organization's history.^[1] For actual processes, this means that experience becomes usable memory content only when it is not merely available but can be reliably connected to a later operational question.

A process memory is therefore neither a mere archive nor an automatic decision-making authority. It creates a controlled space for remembering. Within it, people can identify what happened under which conditions, how a comparable case was handled, and which assessment subsequently remained valid.

Example: A similar case under an older requirement

In an illustrative case, a specialist investigates a deviation in batch C41. They search for earlier operations with the same process phase and a similar material state. Retrieval returns C18: after a long wait, rework was performed, and the subsequent result fell within the intended range.

The case view also shows two differences. C18 was processed using different equipment, and the SOP version in force at the time has since been replaced. In addition, the suspected cause was assessed only as an open assumption. The successful result therefore demonstrates neither that cause nor the permissibility of the same rework for C41.

The specialist can now specifically consult the reviews conducted at the time and the changed requirement. If evidence for today's use is missing, it remains an open prerequisite. The retrieved case provides a concrete direction for investigation, not a ready-made work instruction.

If C18 is reassessed later, this addition should also be identifiable at the next retrieval. The earlier assessment remains historically preserved, while its current assessment status limits further use.

Retrieval means more than full-text search

A search interface can find documents by words. This is often insufficient for a process memory. Comparable cases may be worded differently, while operations with the same name may have entirely different meanings under other conditions. Retrieval must therefore account for domain characteristics and relationships.

W3C SKOS makes it possible to assign preferred, alternative, and hidden labels to concepts and express hierarchical and associative relationships between them.^[5] Such controlled concept systems can help users find suitable cases even when they use a different permissible label. They do not, however, replace substantive case logic.

Good retrieval starts with the current situation. For example, the system can use product type, process step, material state, equipment, deviation class, and applicable rule version. Cases are then displayed not only by similarity but also with their differences: the same cause but different material; the same measure but a different site; a similar measurement but a specification that has since changed.

These differences are not search noise. They are an essential part of the decision aid. A process memory should not create the impression that an earlier case is automatically transferable. It must make visible the conditions under which a comparison holds—or does not.

A process case that can be recalled needs a clear structure

For an earlier operation to remain understandable later, it must be captured as a coherent case. A useful case structure can be divided here into five layers:

• Initial situation: Product, batch, material, site, process phase, environmental conditions, and trigger;
• applicable framework: SOP, specification, rule, model version, and responsible roles applied;
• actual history: Tasks performed, measurements, state changes, deviations, reviews, and rework;
• decisions: Selected measure, alternatives, reasoning, responsible person, and required approvals;
• assessment: Observed effect, later review finding, limitations, and status of the experience gained.

These layers must not merge into a freely written case story. Structured relationships make it possible to search specifically for individual conditions later and reveal differences between cases. Descriptive text nevertheless remains important when a situation cannot be fully expressed through fields and rules.

The context model supplies the framework of meaning. It describes which operational properties shape a situation and how objects, tasks, roles, and conditions relate. The process memory uses this framework to place specific past cases within it. Model and memory are thus closely connected but not identical: the model describes possible meanings; the recalled case documents an execution that was actually observed.

From observed history to approved experience

An event does not become operational knowledge simply because it happened. A substantive assessment lies between observation and reuse. Every potential memory item therefore needs a traceable status, such as:

• observed: The history is documented but has not yet been conclusively assessed;
• under review: Cause, effect, or transferability are being substantively investigated;
• confirmed: The experience has been reviewed and approved for a defined scope;
• restricted: The statement is usable only under explicitly named conditions;
• disproven or replaced: An earlier assumption should no longer serve as valid guidance but remains historically visible.

The assessment should be stored separately from the original history. The history documents what happened. The assessment describes which substantive conclusions may be drawn. This separation prevents an interpretation from retrospectively appearing to be a directly measured fact.

Operational knowledge models place approved knowledge into a usable context for tasks and decisions. The process memory supplies specific, traceable cases of experience for this purpose. Not every case, however, becomes a generally applicable rule. Its value often lies precisely in revealing an exception, risk, or previously unclear condition.

Later insights must not overwrite earlier meanings

Processes, products, and requirements change. A case may have been handled correctly at the time yet no longer serve as a comparison under a later rule. The process memory must therefore be able to distinguish the framework of meaning at the time from today's framework.

Versioned knowledge models record which concepts, relationships, and rules applied at a particular time. The recalled case references this specific version. A search today may still find it but must indicate when relevant conditions have changed since then.

Memory content itself can also be versioned. Experience initially limited to one site may be extended after further reviews. A confirmed measure may later be replaced by a better one. This creates a new assessment version; the historical origin remains unchanged and connected to the original process execution.

A “current state of knowledge” is therefore not a single overwritable text version. It arises from approved statements, their scopes, versions, and evidence that remains accessible. Only this makes it possible to trace why a recommendation was appropriate yesterday and differs today.

Identify missing information specifically during retrieval

Case C18 is usable only to the extent that its conditions at the time were recorded. A missing equipment status must not be filled in during today's retrieval using the status of similar equipment. A later recollection must also be distinguished from a contemporaneous record.

Task-oriented organizational memory systems connect different knowledge elements for use in the work context.[3] For the retrieval described here, this means that the original record, substantive explanation, and later assessment remain identifiable as different foundations.

The search may therefore end with an incomplete case. Its benefit then also lies in making a specific evidence gap visible rather than pretending to provide a complete success story.

Provenance and accountability must be preserved

A process memory loses value when statements are separated from their origin. For every memory item, it should be clear which operation it was derived from, which data were used, who performed an assessment, and whether that assessment was later revised.

W3C PROV distinguishes entities, activities, and agents and provides relationships for usage, generation, derivation, and attribution of responsibility.^[4] This model does not prescribe an operational storage architecture. It does, however, illustrate the kinds of relationships required so that experience does not exist as an anonymous statement in the system.

Data provenance answers where data and statements come from and which operations produced them. The process memory goes a step further: it organizes selected past cases for later situation-specific retrieval. Provenance remains a prerequisite. Recorded provenance relationships support reviewing condensed experience against the original history; missing references must continue to be identified.

A later correction also belongs to memory. If a cause is reassessed or a measure rejected, the earlier assessment must not silently disappear. The case receives a new assessment with date, reasoning, and responsibility. This keeps it visible what counted as knowledge at which time and why that classification changed.

Stored data are not yet memory

An organization can store documents, measurements, task logs, and approvals for years without forming a functioning process memory. A large volume of data does not automatically answer which earlier experience is relevant to the current situation. Nor does it explain whether an earlier approach was successful, only provisionally accepted, or later expressly rejected.

The difference lies in the relationships. An isolated measurement has little value for recall. But connecting it to batch, material, equipment, site, process phase, SOP version, operator's role, deviation, decision, and later result creates a reconstructable case. Only this structure allows targeted retrieval: Which comparable situations existed? Which differences are decisive? Which measures were reviewed? What can actually be inferred from the earlier history?

According to Stein and Zwass, information systems are a vital part of organizational memory; their model describes functions for acquiring, retaining, maintaining, searching for, and retrieving information.^[2] A part, however, is not the whole: practical experience also remains embedded in people, routines, roles, and ways of working. A digital process memory makes selected parts of it traceably accessible; it does not replace these carriers completely.

Distinguish process memory from related concepts

• Document archive: Retains files and records; it does not necessarily organize specific experience for situation-specific retrieval.
• Audit trail: Logs actions and changes; it does not automatically assess which experience follows for later processes.
• Process memory: Connects selected earlier histories to context, substantive assessment, scope, and reuse.
• Knowledge model: Describes concepts, relationships, and rules; a process memory contains specific historical cases within this framework of meaning.
• Process mining: Analyzes event data across many histories; process memory makes assessed experience accessible for later situations.
• Decision support: Prepares information for a current decision; a process memory may be a source but does not make the decision itself.

These concepts can complement one another but must not be equated. A complete audit trail can provide a good factual basis. Only selection, contextualization, and substantive assessment turn it into a reusable case of experience.

What a process memory cannot accomplish

• It does not prove causality. Recurring commonalities may indicate a relationship but do not replace a substantive investigation.
• It does not replace people's practical experience. Knowledge that cannot be fully articulated remains dependent on individual and shared practice.
• It does not justify unlimited data collection. The scope of capture, retention, and access must be appropriate to the operational purpose.

Regulatory finding

Finding: The FDA found that an investigation into a mold finding dated June 4, 2025, failed to account for a previous finding on the same line. The letter identifies the June 2025 finding as Chaetomium cruentum/globosum. The investigation stated that mold had never previously been detected in the Grade A area of that filling line. The FDA instead pointed to Chaetomium globosum recovered on March 20, 2024, on the same line, describing it as the same mold.^[6]

Assessment: A stored history supports an investigation only when relevant earlier findings are considered with their relationships and differences. The responsible people assess their comparability and significance. This CGMP finding requires neither automatic retrieval nor a particular software concept.

Review question: Before an investigation is closed, are earlier findings from the same line or involving the same organism presented with their context and assessed by the responsible people?

Letter dated 27 August 2026 · Source checked on 18 September 2026.

This section presents the selected regulatory finding as stated at the time of the letter. Company responses and subsequent developments are not assessed here; this account does not describe the company’s current compliance status.

C18 is a comparison case, not a release for C41

In the example, different equipment, a new SOP version, and an open causal assumption limit transferability. These very differences make the result useful: the specialist knows which prerequisites must be investigated before a new decision.

A process memory should therefore do more than find matching cases. It also keeps their assessments at the time, later revisions, and current limits accessible. Memory supports a new review when the old case remains visible with its actual status.