Can memory reconsider its own evidence?
ERM’s relational-memory prototype evaluates an event’s pattern alongside declarations about that same event and their sources. This portal brings together the mechanism, dated synthetic checks, reproduction record and open limits.
The research question
If an event’s relations support temporary use while an explicit denial concerns the same event, what should memory show? ERM investigates keeping pattern evidence and declaration evidence distinct rather than hiding that tension inside one label. Finding records matters, but so do the frame and evidence under which they are used.
- The same event frame.
- Event, participants, object, time and source context jointly constrain the inquiry. Another transfer cannot supply missing relations.
- Separate pattern and declaration evidence.
- Show the relational structure and explicit statement together. Neither automatically outranks the other; opposing support in the same frame is reported as a conflict.
- A judgment is not a new fact.
- Reassess support when sources or revisions change. A query result is not inserted into persistent memory as a new fact.
How memory evaluates
The path of a query
Select the target transfer
Identify the event, participants, object, time and source frame. A better-matching event does not replace the target.
Collect live relational support
Evaluate relations visible in the selected revision with current source support. A positive declaration cannot manufacture missing relations.
Compare the declaration in the same frame
Examine explicit support or denial separately. Do not randomly select a transfer when its reference is ambiguous.
Show the judgment with its evidence
Expose support, conflict, missing information or reference ambiguity. Reassess after source withdrawal; querying must not mutate persistent state.
This summarizes the published account of v0.12, with one manually defined meaning bridge. It does not claim automatic extraction of relations from natural language or general meaning learning.
Four cases for reading the mechanism
These summarize recorded behaviors for explanation. No new experiment was run for this portal.
Pattern and counter-declaration
Temporary-use structure is present, together with an explicit denial about the same transfer.
Preserve both evidence sets and expose the conflicting result.
A declaration with missing relations
A positive statement is present, but the required relational structure is incomplete.
Show declaration support without inventing the missing relations.
A supporting source is withdrawn
The source carrying one premise of the judgment is no longer visible.
Reconsider support in the selected view rather than automatically retaining the old answer.
An ambiguous transfer reference
Two transfers fit the same declaration frame.
Report ambiguity instead of choosing one at random.
Experiments and recorded results
The v0.4 validation record dated 6 October 2026 reports preservation of earlier behavior and new pattern–declaration queries in separate groups. These are dated results from developer-authored synthetic checks; they were not rerun here.
| Control group | Recorded matches | What is checked? |
|---|---|---|
| 40 earlier questions | 40/40 | Preservation of earlier detailed answers and evidence. |
| 36 earlier counterexample queries | 36/36 | Preservation of answers and evidence on counterexamples. |
| 25 staged judgments in 20 new scenarios | 25/25 | Pattern–declaration consistency under changing sources and revisions. |
Do not pool these into a single general success rate. Expected judgments were written by the developer; they are not independent human-meaning evaluation.
Technical checks in the same report
- Engine tests
- 348
- Protocol tests
- 16
- Technical checks
- 127
348, 16 and 127 refer to different checking layers. They establish neither validated human meaning nor generalization.
Related writing
Removing one relation
A comparative experiment outline for ERM, rather than a new result.
Making remembering part of the medium
An entry into thinking media with Andy Matuschak and Michael Nielsen.
A reproduction record in ERM
Inputs, file identities, evidence and the limits of human evaluation through ERM v0.12’s dated validation records.
Does memory store records or relationships?
Through ERM: the record of an event, the relationships that make it meaningful, conflict, and changes in source support.
Reproduction and sources
The same answer, supported by the same evidence
Fix the version and inputs
Engine sources and four JSON inputs are identified with SHA-256 hashes. A changed file is a changed examination.
Compare more than labels
Check that evidence belongs to the selected revision, live sources and target transfer as well as matching the final label.
Separate reading from writing
Compare persistent state before and after queries; inspect detailed answers after saving and reloading.
Give a new run its own record
Reproduction in another environment needs its own version, environment, command and results. A file-integrity check is not an engine rerun.
Source register
There is not yet a verified public download link for the original source package. The entries below cite dated reviews already published on this site; local paths and guessed repository links are not published.
Relational-memory design record
Event framing, the pattern/declaration distinction and the manually defined meaning bridge.
Does memory store records or relationships?Pattern–Declaration Consistency protocol
Synthetic scenarios, separate expected results and evaluation rules.
A reproduction record in ERMValidation and comparison report
The 40/40, 36/36 and 25/25 groups and distinct technical checking layers.
A reproduction record in ERMLock and file-integrity review
The published note reports that four input hashes matched the lock; this was not experiment reproduction.
A reproduction record in ERMExisting project text and dated sources
The original editable research entry is preserved below. This portal expands the dated account without publishing a new experiment or prototype source version.
When a pattern and a statement conflict
What should memory show when an event's relational structure conflicts with an explicit statement about it? This ERM entry covers the prototype combining patterns and statements. The RGM / Relational Space H1a experiments have a separate research entry.
Retaining both kinds of evidence
Version 0.12 reports conflict when a temporary-use pattern is supported but explicitly denied within the same transfer frame. Both kinds of evidence remain visible; no source automatically outranks another. The event, participants, object, time and source context constrain evaluation. Another transfer cannot supply missing parts.
Support is reassessed when a source is withdrawn. A positive statement does not create missing relations, and a derived judgment is not stored as a new fact.
Measured scope
The 6 October validation report records preservation of 40 earlier questions and 36 pattern queries, with 25 staged judgments matching expectations across 20 new synthetic scenarios. These are version-specific reported results; tests were not rerun for this entry.
Open boundary
Relations and the meaning link are manually defined. Natural-language parsing, general meaning-equivalence learning and independent human meaning evaluation are not established. The next question is how to test the mechanism with natural inputs and independent assessments.
Source note
Sources reviewed on 14 October 2026: relational memory v0.12 README and pattern–statement validation v0.4 report (6 October). RGM retrieval experiments are not evidence of this prototype's success.
Status and open limits
Prepared on 22 October 2026 by bringing together the existing ERM research entry and two published essays. The described version is v0.12; the research entry’s source cutoff is 14 October and the technical report is dated 6 October. Those dated results are not presented as a new reproduction run.
The meaning bridge and relations are manually specified. Natural-language parsing, general meaning-equivalence learning and independent human evaluation are incomplete. The record states human_meaning_validated=false. Keeping expectations separate from engine answers matters, but does not itself validate human meaning.
Passing synthetic checks ≠ validating human meaning. Matching file hashes ≠ reproducing the experiment. Exposing a conflict ≠ resolving it.
RGM — Relational Space / H1a
RGM / Bağ Uzayı H1a studies evidence retrieval through derived contexts. It is kept separate; retrieval results are not evidence for ERM’s pattern–declaration mechanism.