Compiled Memory vs Governed State: Karpathy's LLM Wiki and the Reflexive Laboratory

Description

Abstract

This paper compares two persistent LLM knowledge architectures: Andrej Karpathy’s LLM Wiki and the Reflexive Laboratory. Karpathy’s system compiles immutable raw sources into a maintained markdown wiki through ingest, query, and lint operations, creating a persistent derivative memory layer that reduces future rediscovery cost. The Reflexive Laboratory preserves transcripts, typed artifacts, registries, canonicality fields, and snapshots in order to govern not only accumulated knowledge but also current authority and admissible continuation. The paper’s core claim is that these systems preserve different objects. Karpathy’s architecture is a compiled-memory system; the Reflexive Laboratory is a governed-state system. To formalize the difference, the paper uses two tools already present in the Reflexive Laboratory series: the locus-of-plasticity framework and the distinction between execution sufficiency and state sufficiency. It argues that both systems relocate adaptive burden away from model weights, but to different external layers: Karpathy primarily to architecture, the Reflexive Laboratory primarily to governance. It further argues that replayability and resumability separate. The paper concludes that wiki systems naturally yield graph candidates, while the Reflexive Laboratory is structurally closer to a governed scientific knowledge graph because it preserves explicit lineage, authority, and validation-bearing state.

Other

This package presents a structured research object from the Reflexive Laboratory program, a transcript-based approach to AI-assisted scientific inquiry. Rather than treating research as a sequence of isolated papers, this work captures both the final manuscript and the underlying process through which it was developed.

The package includes the publication-ready paper alongside supporting materials that document how ideas were formed, revised, and stabilized over time. These materials reflect a “transcript-sufficient” research model, in which key steps in reasoning, drafting, and artifact construction are preserved as part of the research record.

Across the Reflexive Laboratory series, this approach enables:

Process visibility: showing how a research object evolves from early concepts to final form

Artifact traceability: linking claims, figures, and structures to their development history

Reproducibility support: providing enough context to reconstruct how results were produced

AI transparency: making visible how human and AI contributions interact within the workflow

This package should be understood as more than a standalone paper. It is a composite research object that combines:

a formal scientific or methodological contribution

a structured record of its production

supporting artifacts that enable inspection, reuse, and extension

The goal is to demonstrate a different model of scientific publication—one in which the research process itself becomes partially observable, inspectable, and reusable, rather than hidden behind the final text.

This work is part of an ongoing series exploring how AI-assisted systems can support rigorous, transparent, and governance-aware research production.