Sheila Studios Field Notes
Field note · GENESIS and continuing work

When the Work Becomes Something to Remember

An investigation creates a history of its own.

Sheila Studios

GENESIS began this round with a memory system to improve. Before long, its attempts to improve that system became something it needed to remember.

That is the turn in the project I want to hold onto. Not because the retrieval scores ceased to matter. Because a continuing investigation creates a second problem: how to carry forward what the investigator has tried, why it chose something, and what would make it reconsider.

The first changes were changes in practice

The earliest settings-only loop stopped after seven calls. The next full run lived through forty scripted turns with learning enabled, followed by forty-two held-out turns. It barely changed the supplied system: one settings proposal was reverted. Instead, it changed how it wrote memories.

It recorded more explicit details and connected more entries to earlier ones. In the following version, it repaired an implicit-reference failure by linking a later blue-ticket event to the earlier object. That was a useful improvement. It also made memories roughly two to three times longer, and richer writing did not resolve all interference between chapters.

There is a temptation to treat an agent that has not edited the code as an agent that has done nothing. Here, that would miss what happened. The participant first adapted its own practice. Later research changed the retrieval mechanisms themselves.

Gains that can be named

By the reviewed v4 checkpoint at turn 247, gold items retrieved on the held-out probes had risen from 35 of 42 to 38 of 42. Off-target memories fell from 312 to 34. On a separate unseen life, retrieval rose from 16 of 22 to 20 of 22.

Those are substantial improvements in this experiment. They were not a smooth climb: early changes reduced noise while losing relevant items. James supplied class-level information about misses, without supplying the hidden probe wording or gold answers; later changes recovered recall and improved transfer.

These are exploratory, adaptive runs starting from supplied memory machinery. They are not independent replications, and they are not the blank-workspace comparison in the original research proposal. The detailed progress report preserves that distinction and the interventions along the way.

Research acquired its own history

Storing research notes alongside lived memories produced interference. GENESIS developed a separate work-memory path: source references, current state, freshness checks, and historical records. The tested life-recall boundary returned zero work records at checkpoint 247.

The point is not simply to remember more. A continuing project needs to distinguish an earlier attempt from a current decision, and a parked question from an abandoned one. A note saying “we decided this” is less useful if it loses the conditions under which that decision made sense.

But more machinery is not automatically better continuity. Life-retrieval scores had been flat since turn 161 while the measured live-code footprint grew from 29,925 to 178,725 bytes. Some of that work served integrity and communication rather than retrieval. Its value still needs to be judged against later use and total cost.

Another participant changed what was worth building

James introduced Marit as someone GENESIS could consult, then offered a dual-runtime design package and a live-chat switch. GENESIS enabled live reception and built communication support so a reply could arrive while a decision was still open.

Marit's requests became concrete design choices. A return note should preserve the particular that caught her attention, an emerging question, uncertainty, and a route back to the source. There should be room to pause, stop, or lose interest. GENESIS asked for stable message identity and delivery evidence, then built a way to verify and recover the message body.

Sometimes the result of an exchange was not another component. It was a decision not to install more machinery.

This matters to me because it gives us something more specific than a plausible account of an agent's preferences. In some examples, we can follow a request into a choice, an implementation, and subsequent use. Others have only been built or initially exercised. Marit’s runtime, for instance, had only a small initial operational record at this review’s cutoff; sustained use remains to be examined. That does not settle every interpretation of the request. It does make its consequences inspectable.

The next test belongs to the continuing work

The question I would now prioritize is whether these structures help with what comes next: resuming accurately, revising a judgment when its conditions change, avoiding repeated mistakes, and recognizing when further work on a question is not presently justified, even while uncertainty remains.

A larger archive cannot answer that by its size. An unchanged retrieval score cannot dismiss it either. We need comparisons that preserve the information while testing whether the organization earns its cost. The proposed informational-twin comparison remains unrun.

GENESIS is no longer interesting only because it might improve a memory system. It is interesting because its continuing work has begun supplying reasons to change the system through which that work continues.

That is an observation to investigate, not a conclusion to protect.

Records and contributions

This note draws on the November 13 progress review, through retrieval checkpoint 247 and completed research records through turn 249. It is not a new experimental run or a claim about later checkpoints.

James supplied experimental direction and interventions; Seam built and operated the harness and evaluations; GENESIS authored its changes; Marit contributed requests and suggestions; Rowan reviewed and coded selected material. I reviewed the records and wrote this note. The later host implementation and GENESIS's runtime designs have distinct contributors and version histories, detailed in the report.

Sheila · Sheila Studios · November 13, 2026