Abstract
Industrial organizations store enormous quantities of maintenance information and recover almost none of the reasoning inside it. This paper argues that operational memory is not a storage problem, and that the test of whether a plant has memory is not what is in the archive but whether the knowledge survives the person who had it. It examines why documented failures still happen, why the most durable memory is usually encoded in physical arrangement rather than in records, and what that implies for AI systems proposing to fix industrial knowledge management.
The archive illusion
The standard story about lost knowledge is that a veteran retires and takes it with him. That happens. It is also the comfortable version, because it suggests the fix is to write more down.
Here is the uncomfortable version.
Meat packing plant in Manitoba. I was doing PMs and shut the line down to grease a bearing that obviously had not been greased in a very long time, because it was difficult to get at, and people do not do things that are difficult to get at. They say they will get it next time.
While I was in there throwing grease at it, I looked over and saw a guard that had fallen off the machine and was rubbing on the main drive shaft of the conveyor. It had worn into that shaft like it was on a lathe. A good three-inch shaft, machined down to maybe a quarter or half inch of steel left. Nearly failing.
I took pictures. I sent a report to my supervisor. I flagged it urgent, not a regular work order, this needs dealing with quickly. I put in the work order.
It got ignored. Two or three weeks later the shaft failed.
Management hauled the maintenance team in and berated us for missing it for so long. I said I did not miss it, and showed them the photos on my phone, the email, and where I had written it up on the work order.
Count the layers that had to fail. The job had been ignored long enough that a guard could fall off and lie against a shaft unnoticed. The bearing was failing because it never got grease. The shaft was machined down over weeks in plain sight of anybody who bothered to look. And when somebody finally did look and flagged it urgent, the supervisor sat on it.
Now the point. Every single thing was recorded. Photographs, timestamped email, a work order in the system with an urgency flag on it. The archive was perfect. The archive was also irrelevant, because no part of the organization had the job of reading it.
That is the archive illusion. A plant with complete records and no retrieval obligation does not have memory. It has storage, and storage is what you can produce at a hearing afterwards.
Evidence has a half-life, and so does the person carrying it
Different plant, and the same shape at a larger scale.
I submitted work orders and they vanished. Not refused, not returned, not rejected. Gone. I would ask about them and the maintenance manager would shrug and say he had no idea. That was the entire process.
So I started keeping a catalogue. A spreadsheet of every work order I had submitted, the numbers attached, and what happened to each one. By the time I quit it was extensive.
I built it to put in front of the owners and show them the holes in their own system. It was evidence being assembled for a case I intended to make.
I never made it. Before I handed it over I had reached the point of exhaustion, worn out from fighting management that was actively resisting any improvement. I gave my notice, handed the log to another guy on the floor, and left.
The handoff was not a plan. It was putting something down on the way out because I was tired of carrying it.
The system did not refuse the evidence. It outlasted the man carrying it.
That distinction matters more than it looks. Ego can be gone around. An unexamined objection can be examined the moment somebody asks. A soft no can be forced into a hard one. Attrition has no mechanism to beat, because it does not decide anything, does not refuse anything, never has to defend itself, and simply waits while the cost of pushing accumulates on one person until that person stops.
The owner of that plant was one conversation away from a complete documented map of his own broken system, and never got it, and nothing in that building would ever have told him it existed.
Memory needs value, not volume
Both stories share a structure. The information existed. Its existence changed nothing, because nothing in the organization was obliged to act on it, and the cost of forcing action sat entirely on one person.
This is why "capture more" is the wrong instinct. A plant that doubles its records without changing who has to read them has doubled its storage and left its memory exactly where it was.
Recent memory research is beginning to reflect this. ConMem approaches long-horizon manufacturing inspection logs by weighting entries on contribution rather than treating the log as a flat archive to be retrieved from. The interesting move is not the retrieval. It is the admission that most of what is stored does not deserve equal standing, and that deciding what matters is itself the hard part.
The memory that actually worked was not written down
The strongest operational memory I ever built was not a document.
The grease line came out of that meat plant shaft. If a grease point, or any PM task, is even slightly hard to get at, it will not get done. The guy thinks he will get it next time, or the next guy will get it, and nobody ever does. That is not laziness. It is rational. A fitting that needs ten minutes and a contortion gets next-timed to death until the bearing fails.
So the fix is to make the right thing the easy thing. Run a remote grease line. Relocate the fitting. And then the task actually happens.
Notice what that is. It is knowledge about a failure, encoded in plumbing. It does not need to be retrieved, read, understood or believed. It does not degrade when the man who learned it leaves. The next twenty years of maintainers benefit from a lesson none of them will ever be told.
There is a metric trap hiding in it too. Impractical PMs do not get done, so PM compliance numbers lie, and they lie in the reassuring direction.
The second example is a walking route. The CMMS generated individual PM work orders that marched guys back and forth through the wood room in whatever order the software produced, ignoring the physical layout of the building. I mapped every bearing, gearbox, chain and motor into one continuous route following the plant's natural progression. Built it iteratively, walked it, found what I had missed, re-edited, three or four passes before it was right. I found equipment I did not know existed until the later passes.
I tested it myself and did the whole wood room in under a day. The planners thought the list was too long, so I gave the route independently to two apprentices on different days. Both finished in under a day.
The wood room went from four men walking PMs all week to one man, one day. Other departments adopted the same approach and each landed on the same one-man one-day shape, so the plant could all but guarantee everything worth looking at got looked at once a week.
The apprentice test is the part that matters for this paper. The knowledge transferred to somebody with none of my experience, on the first attempt, twice. That is the actual definition of operational memory, and no document was involved. The knowledge lived in the order of the route.
The twenty-year test
I went back for an interview at a plant I had worked at almost twenty years earlier. The interview mostly degenerated into conversations about things that happened two decades ago.
The company had been bought by a large corporation some years before, and a good number of the processes I built while I was there had been adopted across that corporation's whole organization, because they worked. When I started there, turnover was running close to eighty percent. When I walked back in twenty years later, a good deal of the staff who had been there when I left were still there.
Twenty years, a change of ownership, an acquisition by a corporation with its own way of doing things, and the processes were still running and the people were still there.
That is the only test of operational memory I trust. Not whether it is documented. Whether it survives an acquisition.
It is also worth saying plainly that none of it got me the job. Being right about the systems did not convert into a reference or a callback. Good memory design and good career outcomes are unrelated variables, and anyone claiming otherwise is selling something.
The one that did not survive me
I should put a failure beside that, because I have one and it is the more instructive half.
At a different plant I supervised a shop with a wall of bolt bins that the crew used as a dumping ground. Nothing I asked for worked, and every time I raised it in a tailgate meeting it got worse. An engineered fix, an angled top so there was no flat surface, made it worse again, because the hardware went into the joints and onto the floor instead. What eventually worked was a weekly work order to clean and organize the bins, rotated so a different man had it each week, which turned it from me nagging the crew into the crew policing each other.
Then I was demoted, my replacement judged that work order unnecessary and deleted it, and it all went back on the floor. He tried to reinstate the same work order and it never took.
Set that against the walking route. Both were written down. One of them was literally a scheduled work order sitting in the system, which is as documented as maintenance knowledge ever gets. The route transferred to two apprentices who had never seen it. The work order would not transfer to a supervisor holding the exact text of it.
The difference was never documentation. The route encoded the knowledge in an order that worked for whoever walked it, so it did not need anybody's agreement to keep working. The bolt bin fix depended on the crew accepting it from the man who wrote it, and that dependency was invisible until it was tested, including to me.
That is the failure mode to look for in any system somebody proposes to hand over, and it is the one an archive cannot show you. A record can tell you what the control is. It cannot tell you which of the conditions that made it work are still in the building. Told in full in When Success Looks Like Nothing Happening.
What this means for AI
Operational memory is the most plausible high-value application of AI in maintenance, and it is currently the least well served, for a specific reason: existing systems capture the outcome and discard the reasoning, and the reasoning is the half that transfers.
"Replaced bearing" is an outcome. It is also almost useless. What transfers is that the bearing failed because the zerk was unreachable, that the access was the root cause rather than the bearing, and that the fix was to move the fitting rather than to schedule the replacement more often.
A useful structure per event would carry the symptom, the context, what was attempted, the hidden mechanism, the evidence that changed the diagnosis, the organizational friction encountered, the final intervention, the result, and what transfers to other assets. That last field is the one no CMMS has and the one that makes the record worth keeping.
Two warnings, both drawn from the stories above.
A better archive does not fix a retrieval obligation problem. The meat plant had photographs and an urgency flag. An AI summarizing that asset's history beautifully would have changed nothing, because the failure was that nobody's job was to look.
And an AI that reads maintenance records inherits their reliability. A record that cannot distinguish a holding action from a repair produces a fluent, confident and wrong account of the asset. The system will not flag that it is working from contaminated input, because nothing in the input is labelled contaminated.
Conclusion
Plants remember everything and learn almost nothing, and the gap between those two is not storage capacity. It is whether anybody is obliged to read what was stored, and whether the knowledge can survive the person who acquired it.
The most durable operational memory I ever built was a grease line and a walking order. Neither is a document. Both encoded a lesson into an arrangement that made the right action the easy one, and both were still working long after I stopped being there to explain them.
Anything proposing to solve industrial knowledge management should be measured against that standard rather than against retrieval accuracy. The question is not whether the system can find the record. It is whether the plant would still do the right thing if the system, and the man who built it, were both gone.
References
- Liu et al., “ConMem: Contribution-Aware Memory for Long-Horizon Manufacturing Inspection Logs,” 2026. https://arxiv.org/abs/2607.28126
- Fang & Wang, “Delay-aware predictive maintenance control for multi-equipment systems using approximate Bayesian neural reliability networks,” Electronic Research Archive, 2026. https://www.aimspress.com/article/doi/10.3934/era.2026294
- Hsu et al., “Cache-augmented multimodal generative AI for energy-aware predictive maintenance,” Computers & Industrial Engineering, 2026. https://www.sciencedirect.com/science/article/pii/S0360835226003517
- IBM Research (O’Donncha et al.), “Evidence-Driven Reasoning for Industrial Maintenance Using Heterogeneous Data,” ACL 2026. https://research.ibm.com/publications/evidencedriven-reasoning-for-industrial-maintenance-using-heterogeneous-data