Institutional Memory: The Rules Nobody Wrote Down | HL Hunt
Institutional Memory: The Rules Nobody Wrote Down
A written policy covers the cases someone thought to write down. The operating policy includes several hundred further rules about which exception applies, which system is authoritative, and what happened last time someone skipped a step. Those live in people, and every departure removes some without anyone knowing which. At ordinary turnover the staff body has a 3.5-year half-life, and within four years the holder of 55% of the undocumented rules has gone. What leaves first is not the rule but the reason for it.
What you'll learn
The core thesis
An institution's written policy is not its operating policy. The written version covers the cases someone thought to write down. The operating version includes several hundred additional rules about which exception applies when, which system is authoritative, why a particular check exists, and what happened the last time someone skipped it.
Those rules are held in people. Every departure removes some of them, and nobody can say which ones, because the only record of their existence was the person applying them.
Per our exception analysis, the undocumented layer is where a large share of outcomes is actually decided. This article is about what happens to that layer over time, which is that it leaks at a rate nobody measures and is replaced by reinvention.
Our position is that institutional memory is a depreciating asset with a measurable depreciation rate, and that treating it as a permanent feature of an organization is the most common source of repeated operational failures.
The depreciation rate
Take an operation with 34 staff and annual turnover of 18%, which is unremarkable:
- Departures a year: about 6
- Share of today's staff still present in three years: 55%
- In five years: 37%
- Half-life of the staff body: 3.5 years
Now suppose the operation runs on roughly 240 working rules, of which 35% are not written anywhere:
The figure that matters is not the turnover rate but the interaction. An 18% turnover rate sounds like a people problem and is also an information problem, because it sets the half-life of everything the organization knows but has not recorded. Per our decommissioning analysis, the same dynamic governs model documentation and has the same cause: the window for capturing knowledge closes on a date nobody announces.
What gets lost first
The loss is not uniform. Certain categories of knowledge are systematically the least documented and the most consequential.
| Knowledge type | Documented? | Cost when lost |
|---|---|---|
| What the rule is | Usually yes | Low; it is written down |
| Why the rule exists | Almost never | High; a rule without a rationale gets removed by the next efficiency review |
| Which exceptions are legitimate | No | High; per our exception analysis, this is where outcomes are decided |
| Which system is authoritative | Rarely | High; per our accuracy analysis, most data errors trace to this |
| What failed last time | No | Highest; the organization repeats it |
| Who to call about X | No | Moderate and constant |
The second row is the one that produces the recurring failures. A control whose purpose nobody can state looks like unnecessary friction, and the natural response to unnecessary friction is to remove it. The control is then removed, the problem it prevented recurs, and a new control is added, which in a few years will also have lost its rationale.
The fifth row is the reason that cycle repeats rather than resolving. Per our testing analysis, an organization that recorded its failures would be learning from them. One that holds them only in the memory of whoever was present is relearning them on a schedule set by turnover.
How to tell it is happening
The symptoms are distinctive once you know what they are, and they are usually attributed to other causes.
- A recurring incident with a different root cause each time. Often the same root cause, rediscovered.
- Controls being removed as redundant, then reinstated after an event.
- "We used to do it differently and I do not know why it changed."
- One person being the only route to a routine task. Per our segregation analysis, this is also a control weakness for a separate reason.
- New staff producing better questions than anyone can answer. A good diagnostic, because they are encountering the undocumented layer directly.
- Projects reaching the discovery phase and stalling, because the current state cannot be established.
The fifth bullet is worth treating as a measurement instrument. Per our legibility analysis, every question a new joiner asks that nobody can answer identifies a specific piece of undocumented knowledge, and collecting those questions for three months produces a better inventory of the gap than any audit.
A falsifiable claim
Operational incident rates should correlate with tenure composition rather than with headcount or volume. Any operation with incident records and staffing history can test it: regress incident rate on average team tenure, controlling for volume and headcount. If tenure explains nothing, the knowledge in question is adequately documented and the concern here is overstated. We expect a clear negative relationship, concentrated in incident categories involving exceptions and system-of-record questions rather than in categories governed by written procedure, and that distinction is what would confirm the mechanism rather than merely the correlation.
The case against writing it all down
The obvious remedy, document everything, fails for reasons worth taking seriously.
- Documentation that is not maintained is worse than none. A procedure describing a system that changed two years ago actively misleads, and per our notice analysis nobody discovers it is wrong until it matters.
- Writing it down takes time from the people who are busiest, because the people holding the knowledge are the ones everything routes through.
- Much of it is genuinely tacit. Judgment about which customers to believe is not reducible to a rule, and attempts to reduce it produce rules that are wrong at the edges.
- Over-documentation invites compliance with the document rather than the purpose, which is its own failure mode.
- Some of the undocumented layer should not survive. Per our exception analysis, an informal practice nobody can justify may be a practice that ought to end rather than be preserved.
The last bullet is the one that reframes the whole problem. The goal is not to preserve the undocumented layer. It is to make a deliberate decision about each part of it, which requires knowing what is in there. An institution that documents a practice and then abolishes it has done better than one that loses it accidentally and reinvents a worse version.
What actually works
Ranked by cost, and the cheap ones are where most of the value is.
- Record the reason, not the rule. A one-line rationale against each control or check. This is the highest-value documentation in any operation and it takes a sentence per item.
- Collect the questions new joiners cannot get answered. Free, continuous, and self-targeting.
- Write an incident record that includes the cause, and read it before the next change in that area, per our audit trail analysis.
- Name the system of record for every contested field. Per our accuracy analysis, this single document prevents a whole category of error.
- Make exceptions leave a coded record. Per our exception analysis, then the undocumented layer documents itself as a by-product of use.
- Run a handover that is a question list, not a document. The departing person answers; the successor writes. Reversing who writes is the whole trick, because the departing person does not know what is surprising.
- Pair on the rare tasks, not the common ones. The annual process done by one person is the exposure, and the daily process is already distributed.
Item six is the cheapest structural improvement available and it is almost always done backwards. A departing employee writing a handover document records what they think is important. A successor interviewing them records what they do not already know, which is the actual gap.
Item one deserves to be stated as the single recommendation if only one is adopted. Per our fixed cost analysis, controls accumulate and are expensive, and the reason they are never pruned safely is that nobody knows which ones are load-bearing. A rationale column converts that from an unanswerable question into a readable one.
Scenarios and what we are watching
| Scenario | What we would expect |
|---|---|
| Turnover rises across the sector | Operational incident rates rise with a lag of a year or two, and get attributed to volume or to training quality |
| Processes are automated | The rule is captured in code, which is a real gain, and the rationale is captured nowhere, which is the same problem in a new place |
| Exception logging becomes standard | The undocumented layer becomes visible and prunable for the first time |
| Long-tenure staff retire together | A step change rather than a gradual loss, and the most predictable version of this risk |
The second row is the trajectory we would flag. Encoding a rule in software fixes the loss of the rule and makes the loss of the reason worse, because a rule in code is harder to question than a rule in a person. A check that someone could at least explain badly becomes a line of logic nobody can explain at all, and per our decommissioning analysis it then survives every review on the grounds that it must be there for a reason.
Frequently asked questions
At 18% turnover the staff body has a 3.5-year half-life. Of 84 undocumented rules, the holder of about 55% has left within four years.
The reason, not the rule. A control nobody can justify looks like friction, gets removed, and the problem it prevented recurs.
No. Unmaintained documentation misleads, much of the knowledge is tacit, and some undocumented practices should end rather than be preserved.
Record a one-line reason against every control. And reverse the handover: the successor interviews and writes, because the leaver does not know what is surprising.
Key takeaways
- Written policy is not operating policy; a large share of working rules exists only in people.
- At 34 staff and 18% turnover, the staff body has a 3.5-year half-life and 55% of today’s people are gone in three years.
- Of 84 undocumented rules, the holder of about 55% has left within four years, and roughly fifteen leave annually with no record of which.
- The reason for a rule is almost never documented, which is why controls get removed as friction and the problems recur.
- Questions a new joiner cannot get answered are a free, self-targeting inventory of the gap.
- Documenting everything fails: unmaintained documentation misleads, and some undocumented practices should end rather than persist.
- A one-line rationale against each control is the highest-value documentation available and makes pruning safe.
- Encoding rules in software fixes the loss of the rule and worsens the loss of the reason, because code is harder to question than a person.
This analysis is educational and does not constitute operational, compliance, or human resources advice. All figures are stylized illustrations using assumed headcount, turnover rates, and rule counts, and do not describe any particular organization. Turnover and tenure calculations assume a constant annual departure rate applied uniformly, which is a simplification.