Record · research / method
Written · 2026-09-08
By · Clark Glenn Davis
Length · 2,400 words · 3 figures

Provenance as a first-class field in a memory store

A record that cannot say where it came from is not yet a memory. It is a string that happens to be true.

Contents
Entry history
2026-09-08 · written
2026-09-10 · figure 2 corrected
Amendments are appended. Nothing is silently revised.
1 · The failure it fixes

A store fills up quickly, and most of what it holds is defensible on its own terms. The trouble starts at retrieval, when two records disagree and there is no way to tell which one was written by an instrument, which was inferred from a conversation, and which was a guess that has since been repeated often enough to look settled.

The usual reply is to score confidence. That treats a bookkeeping problem as an epistemic one. A record does not need to know how sure it is; it needs to be able to say where it came from, and let the reader do the rest.

Retrieval quality is an editing problem before it is a ranking problem.

2 · Four fields, and no more

Four fields turn out to be sufficient, and a fifth is where the discipline breaks. Origin says where the content came from, verbatim, without paraphrase. Observed says when it was true, which is rarely when it was written. Writer names the process that stored it. Derived_from points at the prior items it was built out of, if any.

origin"transcript:2026-08-30T14:02Z#l88"
observed2026-08-29
writerarbor.ingest v4
derived_from[]
confidenceomitted — never inferred after the fact
Figure 1 — a stored item as written. Captured from the store, 2026-09-08.

The fifth field is confidence, and it is left out on purpose. Confidence assigned at write time is an opinion the store then carries forever; confidence assigned later is a reconstruction dressed as a measurement.

3 · What retrieval does with them

With the four fields present, a conflict stops being a tie. The record whose origin is an instrument reading outranks the record whose origin is a summary of a summary, not because it scored higher but because the chain is visible and short.

Figure 2 — tool calls across 20 sessions, 273 spans
Figure 2 — tool calls across 20 sessions, 273 spans. Garage telemetry, 2026-08-14.
4 · Where it fails

It fails wherever the origin is honest but useless: a long conversation with no stable address, a document that has since moved, a process that no longer exists under that name. Those cases are not rare, and the store keeps them rather than tidying them away. An unresolvable origin is still evidence about how the record was made.

The same rule governs this site. Every claim in the archive carries its source, and where the source cannot currently be verified, the page says so instead of quietly dropping the claim.

Sources
1Store capture, 2026-09-08 · arbor-item-88.json — as written, unedited
2Retrieval run, 400 items, 2026-09-07 · method in the research record, 2026-08-24
2026-09-08 · Method

Provenance as a first-class field in a memory store

A record that cannot say where it came from is not yet a memory. It is a string that happens to be true.

Clark Glenn Davis2,400 words
1 · The failure it fixes

A store fills up quickly, and most of what it holds is defensible on its own terms. The trouble starts at retrieval, when two records disagree and there is no way to tell which was written by an instrument, which was inferred from a conversation, and which was a guess repeated often enough to look settled.

The usual reply is to score confidence. That treats a bookkeeping problem as an epistemic one. A record does not need to know how sure it is; it needs to say where it came from.

Retrieval quality is an editing problem before it is a ranking problem.

2 · Four fields, and no more

Origin says where the content came from, verbatim. Observed says when it was true, which is rarely when it was written. Writer names the process that stored it. Derived_from points at the prior items it was built out of, if any.

origin
"transcript:2026-08-30T14:02Z#l88"
observed2026-08-29
writerarbor.ingest v4
derived_from[]
confidence — omitted
Figure 1 — a stored item as written. Captured 2026-09-08.

The fifth field is confidence, and it is left out on purpose. Assigned at write time it is an opinion the store carries forever; assigned later it is a reconstruction dressed as a measurement.

3 · What retrieval does with them

With the four fields present, a conflict stops being a tie. The record whose origin is an instrument reading outranks the one derived from a summary of a summary — not because it scored higher, but because the chain is visible and short.

Figure 2 — tool calls across 20 sessions, 273 spans
Figure 2 — tool calls across 20 sessions, 273 spans. Garage telemetry, 2026-08-14.
4 · Where it fails

It fails wherever the origin is honest but useless: a conversation with no stable address, a document that has moved, a process that no longer exists under that name. The store keeps those rather than tidying them away.

Sources
1Store capture, 2026-09-08 · arbor-item-88.json — as written, unedited
2Retrieval run, 400 items, 2026-09-07 · method in the research record, 2026-08-24
3Prior entry · Reading a memory store as a document rather than a database
Entry history
2026-09-08 · written
2026-09-10 · figure 2 corrected
Amendments are appended. Nothing is silently revised.
Next
The claims ledger
© 2026 C. G. DavisNot all indexed.