Institutional memory in economic policy is held in definitions, not in people

Institutional memory is usually discussed as though it were recollection: the official who was in the room the last time this was tried, the adviser who remembers how the previous programme was wound up. That is the least of it, and the easiest part to replace. What an economic policy institution has to retain is duller and far more fragile. It has to retain what its numbers mean. The unit a figure is stated in, the base year an index counts from, the vintage of an estimate, the definition that decides who falls inside a category and who falls outside it: none of that lives in anyone's recollection. It lives in a sentence attached to a number, and the sentence is the part that goes missing. A definition has no constituency. Nobody is promoted for defending one, nobody's budget depends on it, and its loss never appears on an agenda until something built on it has already been decided at the wrong size.

Where this is written from

The vantage should be named before anything rests on it. I am an analyst writing inside The Cefa Brief, working from published statistics and public documents and from nothing else. The figures below are annual series published by American statistical agencies, redistributed by FRED and retrieved for this site on 19 September 2026, read alongside each agency's own description of what its series counts. I hold no post in a finance ministry, no seat on any committee, and no sight of a staff file, a handover note or a minute. Everything I can check is a thing somebody has already published.

That position cannot observe forgetting. Forgetting happens inside an organisation, in the gap between one official leaving and a successor arriving, and none of it is published. What the position can observe is the residue. A published sentence that describes a figure wrongly. A paper that quotes an index without its base. A briefing that reaches for a rate regime long since ended as though it were the present one. This piece is therefore built from evidence of one particular kind: errors in public text, read as the visible end of a loss whose middle is not visible from here. Where that inference is carrying the argument I will say so rather than let the numbers appear to do work they cannot do.

The unit is lost before the number is

The clearest case available is one this publication produced itself. Construction spending on manufacturing plant, series TLMFGCONS, is published by the Census Bureau in billions of dollars at a seasonally adjusted annual rate, and its 2025 observation is 215. Because that series states each period at a yearly rate, 215 billion dollars is a figure for a year. It went into print in these pages described as a month's spending, and the description had to be corrected in public. The error runs in one direction only. Read as a month, 215 implies 215 multiplied by twelve, or 2,580 billion dollars, across the year, which is 2.58 trillion. The Census Bureau's own figure for construction of every type put in place in 2025 is 2.24 trillion. Factory building alone would have exceeded all American construction of every kind by 0.34 trillion dollars. The monthly figure the series actually implies is 215 divided by twelve, roughly 17.9 billion. The number in the published sentence was correct throughout. Everything that made it mean something had fallen off in transit.

Then the base, then the regime

A second loss is slower and produces no correction, because nothing in it is false. Industrial production is an index on a 2017 base, where 2017 equals 100. Its readings run 95.1 in 2020, 99.3 in 2021, 101 in 2022, 100.8 in 2023, 100.1 in 2024 and 101.2 in 2025. The latest figure sits 1.2 points above the base, the whole six-year range is 6.1 points, and 2025 stands 0.2 points from 2022. Someone working in 2017 reads 101.2 as a statement about a year they lived through. Someone who arrived in 2023 reads it as a number near a hundred, anchored to a year nine years gone. Neither is wrong. The base has simply stopped being a memory and become a convention, and conventions are the first thing dropped when a figure is summarised for someone senior.

Rates lose their history the same way and at greater cost. The federal funds effective rate reads 0.38 percent in 2020, 0.08 in 2021, 1.68 in 2022, 5.02 in 2023, 5.14 in 2024 and 4.21 in 2025. The climb from the 0.08 low to the 5.14 peak is 5.06 points; the fall since the peak is 0.93. The thirty-year fixed mortgage rate reads 3.11 percent in 2020, 2.96 in 2021, 5.34 in 2022, 6.81 in 2023, 6.72 in 2024 and 6.6 in 2025, so a household borrowing in 2025 is quoted 3.64 points above one borrowing in 2021. Somebody whose whole working life falls after 2021 has worked inside a single regime, and neither rate has returned to its 2021 reading in any year since. For that person the 2021 readings are not history in any usable sense; they are a line in a table. Hurdle rates, appraisal habits and the instinct for what a long-dated project can carry are all formed against a regime, and a regime nobody in the room has worked inside is not available to the room as judgement.

What the loss costs

The cost is not embarrassment. It is decisions taken at the wrong size, and official descriptions the public does not recognise. Consider the difference between a level and a rate of change, which is the most consequential definitional loss in current economic communication. The consumer price index for all urban consumers, on a 1982-84 base of 100, reads 258.9 in 2020, 271 in 2021, 292.6 in 2022, 304.7 in 2023, 313.7 in 2024 and 322 in 2025. The largest single step is 21.6 points, into 2022. The most recent step is 8.3 points, the smallest in the window. An institution holding the rate in memory reports that the problem has receded, and the arithmetic supports it. An institution holding the level reports that the index stands 63.1 points above 2020, which is 24.4 percent higher, and that it has not fallen in any year of the window. Both statements are true. Only the second describes what a household is looking at. Where the first is the only one retained, the resulting communication is accurate and unrecognisable at once, and the space between the two gets filled by somebody with no obligation to be accurate at all.

The opposite failure

Memory also fails by persisting. An institution that carries an assumption long after the thing it described has changed is not remembering; it is holding dead weight that looks like continuity from the outside. The M2 money stock, in trillions of dollars, reads 17.7 in 2020, 20.54 in 2021, 21.61 in 2022, 20.87 in 2023, 21.13 in 2024 and 21.96 in 2025. The step from 2020 to 2021 is 2.84 trillion, the largest in the window by a wide margin. The series then falls once, by 0.74 trillion into 2023, the only decline in six observations, and by 2025 it stands 0.35 trillion above its earlier peak. M2 carries a long reputation as a leading indicator, and that reputation is the assumption at issue. The expansion, the single contraction and the recovery are all in the figures above. What is not in them is any account of what any of it predicted, and this site holds nothing that would supply one. Treating an inherited reputation as though it were a finding is a memory failure of the second kind, and it is harder to see than the first because nothing about it looks like an error.

What preserves it

Not people, on this evidence. People leave, and those who stay remember events rather than definitions. Three things preserve it and all three are clerical, which is why they are hard to fund and easy to drop. The first is the unit string, carried with the number everywhere the number goes and copied rather than paraphrased; the string governing the factory-construction series is seven words long, those seven words would have stopped the error described above, and paraphrasing them instead of copying them is exactly what destroyed them. The second is the retrieval date, because a figure without one has a vintage that must be guessed, and the guess is always the present. The third is the published correction, the only one of the three that puts memory back into an organisation rather than merely holding it, since a correction is a definition restated where people will collide with it. What none of this establishes, from where it is written, is whether any particular institution behaves this way, because nothing published measures an organisation's retention of its own definitions. The claim is narrower and I will hold it there: the losses are visible in text, their cost is arithmetic, and the three remedies are cheap enough that an institution declining them is saying something about what it expects to be held to.