Search finds what was written. Memory notices what a senior colleague would notice.

Firms that have bought enterprise search usually describe the same experience: it works, people use it for a month, and then they stop. Not because the results are bad. Because the results are documents, and the thing they needed was an answer, and turning ten documents into an answer is the work they were trying to avoid.

Search is good at a specific job: you know roughly what was written, and you want the passage. That job exists and search is the right tool for it. What it cannot do is notice.

What can a structured memory raise that search cannot?

Six kinds of thing, each of which is a statement about the record as a whole rather than about any one document in it.

What keeps coming up. Customer concentration named in three documents across four months, by three people who never compared notes. No single mention is remarkable; the recurrence is the finding, and it exists only if something is counting.

What looks like an earlier case. Two portfolio companies that showed this pattern before a covenant reset. This is a similarity between entities, not between strings, and it needs the join across systems before it can be computed at all.

Where the record disagrees. A certificate says one figure and the audited accounts say another. Search will happily return both and say nothing about the fact that they conflict.

What should have happened by now. An approval condition with no evidence of discharge. This one is the clearest illustration of the difference, and it deserves its own heading.

What moved since you last looked. Covenant headroom halved over a year. A movement is a relation between two documents' values; it is not written down in either.

What was expected, against what happened. The base case the committee approved on, beside the accounts.

Why can search never find an absence?

Because a search index is a map from words to documents, and an absence has no words. There is no document containing the sentence “the confirmatory diligence was never completed”. What exists is a March minute recording a condition, and then nothing — and nothing is not a search result.

The most valuable finding in a quarterly review is routinely a document that does not exist. No index can return it.

A structured memory can, because it holds the condition as a typed thing with an expected discharge, and can observe that no later document satisfies it. That is the same move a good analyst makes and the reason it takes them a day: they are not searching, they are reconciling a list of expectations against a corpus.

Is this the system making judgments?

No, and the distinction is deliberate. None of the six is a verdict. Each finding names its cases, names its threshold, and cites its documents. Your firm sets the threshold — three mentions in four months is a parameter, not a law — and the judgment stays with the person reading.

This is not modesty. A system that returned conclusions would have to be trusted, and the buyers this is built for do not trust things they cannot check. A system that returns cited observations with a stated threshold can be argued with, which is the only way it survives contact with a committee.

Why do enterprise search projects stall in month two?

Because the first month is spent confirming it works, and the second is spent discovering that working was never the problem. The tool returns relevant documents. It returned relevant documents before, more slowly. What nobody's job got easier at is the part after the documents arrive.

The honest version is that search moves a cost rather than removing one. Finding the ten documents was perhaps a fifth of the work; reading them, reconciling what they say, noticing what is absent and deciding what it means is the other four fifths, and it is still sitting on the same person's desk.

This is also why retrieval-augmented assistants inherit the problem. Putting a model between the person and the passages automates the reading, which helps — but it does the reading fresh each time, from whatever was retrieved, with no memory of what it concluded yesterday and no way to notice that a document which should exist does not. The improvement is real and it is an improvement to the same fifth.

Doing the work at write time is what changes the shape: read once, write down what was established, and let the question be answered from the structure rather than from the documents.

Where is search still the right tool?

When you know the document exists and want it. When you are looking for a phrase, a name, a form of words in a contract. When the question genuinely is “where did we write that down”.

The two are not competitors so much as different questions. Search asks what was written. Memory asks what the record establishes — including where it contradicts itself and where it is silent. Most firms have bought the first and assumed it would deliver the second, which is why the month-two drop-off is so consistent.

The same six kinds appear in every discipline, in that trade's own vocabulary: a covenant instead of a concentration, a clause position instead of a precedent, a rolled-forward workpaper instead of an unreleased condition. Pick your discipline and the shapes are the same.