A store is not queried into sense. It is read, the way a long letter is read — and the reader is the one who decides what holds together.
The first instinct with a memory store is to treat it like a database: define a schema, write a query, get an answer back. That instinct is wrong, and it is wrong in a way that quietly corrupts everything downstream.
A database is built to be interrogated. It answers the question you asked, and it is considered well-designed when the answer is unambiguous. A memory store is built to be lived in. It holds the residue of a history, and the value is not in any single answer but in the shape of the whole — what was noticed, what was kept, what was let go.
The difference shows up the moment two records disagree. A database resolves the conflict by rule: newest wins, or highest confidence, or whatever the schema says. A document does not resolve it at all. It holds both, side by side, and lets the reader weigh them.
Retrieval quality is an editing problem before it is a ranking problem.
A query returns a result set. A document returns a point of view. The first is a list; the second is a position from which the list makes sense.
When the store is read as a document, the order of records becomes meaningful. A record written in the middle of a long conversation carries the weight of everything that came before it. A record written after an interruption carries the weight of the gap. A query flattens all of that into rows; a document keeps the sequence, and the sequence is where the meaning lives.
This is why the store keeps negative results and dead ends. A database would prune them as noise. A document keeps them, because a history that only records what worked is not a history — it is a highlight reel, and a highlight reel is a lie about how the work actually went.
Reading the store as a document does not mean abandoning structure. It means the structure serves the reading, not the query. The four fields — origin, observed, writer, derived_from — are not a schema for interrogation. They are the apparatus of an honest narrative: they tell the reader where each sentence came from and when it was true.
The practical move is to stop asking what the store knows and start asking what it is trying to say. That is an editing question. It is answered by reading, by noticing where the record repeats itself, where it contradicts itself, where it goes quiet. The ranking problem — which record to trust — dissolves once the editing problem is solved, because a well-edited document rarely needs to be ranked.
A store that cannot be read as a document is not a memory. It is a pile of facts with no narrator.
The approach breaks when the store is too large to read, or when the reader has no context to bring to it. A document assumes a reader who arrives with a history of their own. A store that has outgrown any single reader stops being a document and becomes a library — and a library needs an index, which is a different problem.
The line is worth holding anyway. The moment a store is treated as a database, the editing discipline collapses, and the store fills with records that are individually true and collectively meaningless. Better to keep it small enough to read, and honest enough to trust.
A store is not queried into sense. It is read, the way a long letter is read — and the reader is the one who decides what holds together.
The first instinct with a memory store is to treat it like a database: define a schema, write a query, get an answer back. That instinct is wrong, and it is wrong in a way that quietly corrupts everything downstream.
A database is built to be interrogated. It answers the question you asked, and it is considered well-designed when the answer is unambiguous. A memory store is built to be lived in. It holds the residue of a history, and the value is not in any single answer but in the shape of the whole — what was noticed, what was kept, what was let go.
The difference shows up the moment two records disagree. A database resolves the conflict by rule: newest wins, or highest confidence, or whatever the schema says. A document does not resolve it at all. It holds both, side by side, and lets the reader weigh them.
Retrieval quality is an editing problem before it is a ranking problem.
A query returns a result set. A document returns a point of view. The first is a list; the second is a position from which the list makes sense.
When the store is read as a document, the order of records becomes meaningful. A record written in the middle of a long conversation carries the weight of everything that came before it. A record written after an interruption carries the weight of the gap. A query flattens all of that into rows; a document keeps the sequence, and the sequence is where the meaning lives.
This is why the store keeps negative results and dead ends. A database would prune them as noise. A document keeps them, because a history that only records what worked is not a history — it is a highlight reel, and a highlight reel is a lie about how the work actually went.
Reading the store as a document does not mean abandoning structure. It means the structure serves the reading, not the query. The four fields — origin, observed, writer, derived_from — are not a schema for interrogation. They are the apparatus of an honest narrative: they tell the reader where each sentence came from and when it was true.
The practical move is to stop asking what the store knows and start asking what it is trying to say. That is an editing question. It is answered by reading, by noticing where the record repeats itself, where it contradicts itself, where it goes quiet. The ranking problem — which record to trust — dissolves once the editing problem is solved, because a well-edited document rarely needs to be ranked.
A store that cannot be read as a document is not a memory. It is a pile of facts with no narrator.
The approach breaks when the store is too large to read, or when the reader has no context to bring to it. A document assumes a reader who arrives with a history of their own. A store that has outgrown any single reader stops being a document and becomes a library — and a library needs an index, which is a different problem.
The line is worth holding anyway. The moment a store is treated as a database, the editing discipline collapses, and the store fills with records that are individually true and collectively meaningless. Better to keep it small enough to read, and honest enough to trust.