Does memory store records or relationships?

Memory and meaning — 03

We can record that someone gave an object to someone else. Who gave it, who received it, which object, and when?

Yet this record alone may not tell us what the event meant. The object might have been given as a gift, sold, or made available for use with an obligation to return it.

There is a distance between remembering an event and remembering the relationship that event established.

One of the questions explored in ERM’s relational memory prototype lies in that distance: When retrieving a record, can we also evaluate the connections that make it meaningful?

An object changes hands

Consider a hypothetical example. Deniz gives Ece a camera to use for a while.

If we store this event only as “camera transferred,” some later questions remain open. Was permission to use it given? Was an obligation to return it recorded? Was ownership retained at the time of transfer?

We could also place these details inside a single “loan” label. But another question then arises: which records does the system rely on when using that label?

The approach in ERM attempts to make the grounds for a specific meaning visible through relationships. It evaluates how elements such as the transfer, people, object, permission, and obligation connect.

The relationships themselves are recorded too. Records and relationships are therefore not alternatives to one another. The issue is whether recall is limited to finding an individual record, or can also examine the structure among records.

The frame of a record

Bringing relationships together is not sufficient on its own. The relationships must belong to the right event.

If another camera transfer occurred between Deniz and Ece, its records should not fill gaps in the first event. The same people being present does not establish that it is the same event.

Time and source context are part of this distinction too.

ERM v0.12 targets a specific transfer record. It evaluates the event, people, object, time, and source context through that target. Another transfer does not replace the target because it happens to match better.

This constraint matters. A relational system is defined by which connections it should not make as well as those it can make.

The structure of an event and what is said about it

Let us extend the example.

Suppose the recorded relationships support a temporary-use structure. Yet there is also an explicit declaration in the same transfer frame stating that this is not return-based use.

What should memory show?

This version of ERM does not automatically give one side priority. It reports the result as conflicting, showing structural support and the explicit opposing declaration with their respective evidence.

Two different things are preserved: the pattern formed by the records, and what has explicitly been said about that meaning.

This does not mean the conflict has been resolved. It provides a starting point for examining where the conflict lies.

The opposing declaration must genuinely belong to the same frame. A statement at another time, about another object, or in another source context is not treated as opposition to this transfer.

Remembering without filling in what is missing

Consider the reverse. Some relationships needed to establish the temporary-use structure are missing, but there is a positive declaration in the same frame.

The prototype can show the declaration’s support. It does not generate the missing relationships.

“It was declared to be return-based use” and “the supporting relational structure is complete” remain separate explanations.

Likewise, failing to find a structure does not directly produce a negative conclusion. The absence of a supporting record is not the same as a record supporting the opposite.

Ambiguity of reference is another boundary. If two transfers fit the same declaration frame, the current representation cannot distinguish which one the declaration refers to. The system reports uncertainty instead of choosing one at random.

These positions require assessing the basis of an answer before counting how many answers memory provides.

What changes when a source is withdrawn?

In the previous essay, we asked where old knowledge continues to be used when new knowledge arrives. In ERM, a similar question arises through the support for a judgment.

If the source of the negative declaration causing the conflict is withdrawn and the relational structure remains supported, the result can return to a supported state. If a source supporting the structure is withdrawn, the structure should not continue to be treated as complete.

Furthermore, even if a record remains in general memory, it cannot complete the structure of a selected event revision if it is absent from that revision.

Recall thus moves beyond presenting a previously given answer again. The support still present in the selected view is re-examined.

The judgment arising from this examination is not written into memory as a new fact. The distinction between a query result and the recorded grounds is preserved.

What does this prototype tell us?

The scope is narrow: ERM v0.12 evaluates one manually defined connection between a temporary-use pattern and the meaning of return-based use.

It does not read a natural-language narrative and extract all these relationships itself. It has not learned general equivalences of meaning or explained human memory.

The validation report dated 6 October states that 25 staged judgments across 20 developer-created synthetic scenarios matched the expected results. This is a development check that encoded relationships are evaluated consistently under the stated conditions. It does not replace independent human assessment of meaning.

It nevertheless opens concrete ground for research: we can examine which relationships a meaning judgment rests on, where it conflicts, and when it loses support.

That makes the title’s question difficult to answer by choosing just one side.

Memory can hold records; it can also record relationships. What merits research is how they work together during recall.

When we remember something, do we merely find its record, or also re-evaluate the connections that make it what it is?

---

Source note — 15 October 2026: This article draws on the ERM relational memory v0.12 design document and the Pattern–Declaration Consistency validation v0.4 report. The camera example is illustrative; it does not demonstrate natural-language input to the prototype. No new tests were run for this article.

ERM research page · How does a system know that knowledge has changed?