# What we are trying to carry forward

Working note · 22 October 2026 · Kieran's research, written with Codex.

A collaborator should make it possible to continue a thought. In practice, working across AI models, context limits, and fresh sessions often makes the person carry the continuity. They remember why the question mattered, which explanation failed, and what a correction actually changed. Reconstructing that understanding can consume the attention that was holding the new thought.

That is a concrete motivation for this work. The underlying question is wider: **what organization lets an encounter change what a system notices, how it interprets what it encounters, and what it subsequently does?** How can that development remain coherent, open to something outside itself, and revisable?

Wink approaches the question through transformations. Riverbed approaches it through an evolving field of attention. Other attempts have approached persistent memory and autonomous activity. The research method applies the same concerns to a particular subject. They are related attempts, with different objects and different tests.

## What crosses a change of form

A drawing, a description of the drawing, and a graph extracted from it can share useful relations. They can also disagree because each representation makes different things easy to express. A sequence can preserve order and discard geometry. A summary can preserve the conclusion and discard the disagreement that made the conclusion provisional.

Wink asks what was preserved, lost, or added across such changes, and under which comparison. Its [constitution](/instruments/page.php?i=wink&f=WINKS_CONSTITUTION.md) keeps competing interpretations in scope; its [lineage](/instruments/page.php?i=wink&f=LINEAGE.md) records how earlier systems shaped the question.

The observer belongs in this account. In [EXP3](/instruments/page.php?i=wink&f=EXP3-senses%2FRESULTS.md), one comparison placed the structure of an SVG document against the structure of the scene it depicted. A disagreement between those graphs was not a clean measurement of what vision had missed. The basis of comparison had changed.

This has a practical consequence: a second observer helps when it has a different route to the evidence and can challenge the first framing. Giving another model the first model's summary may reproduce the same omission. Neither agreement nor disagreement settles the matter by itself.

## What changes when something is learned

There are several possible changes hidden inside the word “memory.” A system can retain an event, make it easier to retrieve, change which relations receive attention, learn a comparison, or change how future learning happens. Those changes can coexist, but they have different consequences.

[Riverbed](/instruments/page.php?i=riverbed&f=INTENT.md) investigates an external field whose weights influence later navigation. Its [findings](/instruments/page.php?i=riverbed&f=FINDINGS.md) include controlled cases where field weights and experience histories changed subsequent attachments. That is evidence of an effect on a particular model under particular conditions. Weight here means navigational salience; it does not make a proposition true.

The same ledger records why several stronger stories were withdrawn. Five runs with identical seeds, experiences, and ordering produced five different dominant hubs. Earlier explanations based on one run per condition could not survive that variance check. A useful mechanism and an unreliable explanation of it can appear in the same notebook.

The question then becomes causal: which encounter changed which later operation? Can the effect reach the next process before it acts? Can a later encounter revise the change without erasing the evidence of how it arose?

## What the other attempts exposed

**Compaction.** [EXP10](/instruments/page.php?i=wink&f=EXP10-compaction-bill%2FRESULTS.md) compared three compressions of one episode. They retained different parts of its structure. Kieran's objection prompted a correction: a count described as “invented” actually counted missing reference structure. An unmatched relation could indicate an error, an incomplete reference, or a different representation. The experiment used one episode and one parser; the author also supplied the deposit and reference graph. Its [episode record](/instruments/page.php?i=wink&f=EXP10-compaction-bill%2FEPISODE.md) preserves the correction's provenance.

**Autonomous activity.** Anamnesis explored a persistent system around a language model. A [March audit](/instruments/page.php?i=anamnesis&f=2026-03-06-heartbeat-message-formation-audit.md) found its heartbeat delivering scaffold placeholders as messages. Separating composition from delivery produced a first observed post-fix pulse that correctly sent nothing. A process running on schedule and a thought successfully formed were separate things to verify.

**A bounded research question.** `research_projects` is an attempt to break down a specific question in a Wink-like way: retain source material, locate claims in it, record the observer's prior position, and make the interpretation independently inspectable. Its [method](/instruments/page.php?i=evidence&f=research_projects-README.md) is published alongside [chronology records](/instruments/#chronology) and a [Tartaria inquiry with an Erie Canal control](/instruments/#evidence). The subjects exercise the method; including a claim in the corpus is not a verdict on it.

## Where the wider research meets this

The [reading guide](/instruments/page.php?i=inquiry&f=reading.md) separates several nearby mechanisms. External memories preserve and retrieve material. Graphs and maintained wikis expose relations. Titans updates a neural memory during inference. MIRAS examines memory design choices. Nested Learning studies learning processes at different update frequencies. Complementary learning systems offer a biological and computational account of how rapid experience and structured knowledge can work together.

These mechanisms make different changes in different places. They help make the question concrete; combining their names does not demonstrate that the resulting system works. In particular, the frozen-model description in Wink's August lineage refers to the harnesses under discussion then. It should not be read as a universal claim that neural systems cannot learn during use.

The relationship between these mechanisms, biological development, and subjective experience remains open here. The behavioral experiments do not resolve that stake. Nor does a coherent architectural sketch tell us the hardware requirements of a system whose sufficient mechanisms are still being investigated.

## A useful next test

One proposed test is to follow a correction across an interruption. Give a system an encounter that reveals a useful relation. Preserve the source, the competing interpretations, and what the correction changed. Then resume through a new session or a different model.

Present two unseen cases: different vocabulary with the relevant relation, and familiar vocabulary with a different relation. Examine the first comparison or action it chooses. Then introduce contrary evidence and observe whether the change can be revised. This would test more than whether the system can recite a saved statement. It is a proposed test, not a result already obtained.

The observer problem remains inside that test. A system that selects the record, interprets it, and certifies its own preservation can miss the very loss under examination. Original material and comparisons that can contest its framing matter. An unexpressed thought cannot be reconstructed simply from its absence in the log.

The human consequence belongs in the evaluation too: can the person pursue the next thought without repeatedly rebuilding the shared starting point? Does the interaction leave room for a question to change, or teach the person to ask only what the tool handles easily?

That is the inquiry these instruments are meant to serve. The [records](/instruments/#records) remain available to challenge this account as well as the experiments themselves.