Skip to content

🧠 DSOM vs. LLM WIKI: Comparative Analysis & Adoption Strategy

Entry Point 8: This document serves as the Sovereign Knowledge Entry Point. See START-HERE.md for the master onboarding roadmap.

In honor of Andrej Karpathy's "LLM WIKI" vision, which inspired the formalization of the Ingest, Query, and Lint protocols within the DSOM framework.

Andrej Karpathy's "LLM WIKI" Gist describes a paradigm shift from standard Retrieval-Augmented Generation (RAG) to a persistent, LLM-maintained knowledge base. Instead of parsing raw documents on every query, the LLM continuously reads, synthesizes, and cross-references information into a structured Markdown wiki. The human is the curator and the LLM is the tireless maintainer.

Remarkably, our Deep State of Mind (DSOM) framework already implements the vast majority of this vision. Let's break down the parallels, list our current implementations, and define how we can adopt Karpathy's remaining concepts to elevate our intelligence architecture.

1. What We Already Do (The DSOM Baseline)

Karpathy defines three architectural layers and several core operations. Here is how our existing DSOM protocols map perfectly to his concepts:

LLM WIKI Concept DSOM Implementation Status
The Schema (Instructions on how the LLM should maintain the wiki) AGENTS.md (The Sovereign Constitution) and SKILL.md files that dictate exact behavioural guardrails and operational workflows. Implemented (Highly Mature)
The Wiki (LLM-maintained Markdown files) The Sovereign Markdown Palace (docs/governance/, hall_facts, etc.). Implemented
index.md (Content catalog for navigation) palace_registry.md (Our spatial memory map). Implemented
log.md (Chronological tracking) HISTORY.md (Universal Ledger) and CHANGELOG.md. Implemented
Chronological State Logging checkpoint_summary.txt, task.md, and walkthrough.md. Implemented

In many ways, DSOM is a highly advanced, production-grade implementation of the LLM WIKI concept, specifically tuned for software engineering and DevOps infrastructure.

2. How LLM WIKI Can Make Us Better (The Gaps)

While we have the structure, the LLM WIKI highlights three operational verbs that we can adopt to make the DSOM Palace self-sustaining and compounding: Ingest, Query, and Lint.

A. The "Ingest" Protocol (Automated Synthesis)

The Concept: When a new raw source (e.g., an article, a NOSS PDF, a meeting transcript) is introduced, the AI shouldn't just read it. It should extract the data, summarize it, create new concept pages, and link it into the existing Wiki. Our Adoption: We need a strict separation between raw_sources/ and the synthesized Palace. We should create a dsom-knowledge-ingester skill. When you drop a PDF into raw_sources/, I execute this skill to: 1. Extract the text. 2. Synthesize a Markdown page in the appropriate Palace Hall. 3. Update palace_registry.md. 4. Update SUMMARY.md and mkdocs.yml (using our new Rule 14).

B. The "Lint" Ritual (Automated Health Checks)

The Concept: The AI periodically health-checks the Wiki for contradictions, orphan pages, and stale claims. Our Adoption: As the Sovereign Markdown Palace grows, context can rot. We should introduce a palace-auditor skill (a monthly or ad-hoc ritual) where I crawl docs/ and .agents/ to: 1. Find broken Markdown links. 2. Flag rules in AGENTS.md that contradict each other. 3. Identify "orphan" documents that are not listed in palace_registry.md or SUMMARY.md.

C. The "Query" Loop (Compounding Answers)

The Concept: When the LLM generates a really good answer, comparison, or analysis, it shouldn't disappear into the chat history. It should be filed back into the Wiki. Our Adoption: We currently do this manually when I create an "Artifact" and we save it. We should establish a core behavioural rule: "Any time a complex architectural analysis or troubleshooting guide is generated in chat, the AI must proactively propose saving it as a persistent .md document in the Palace."

3. Proposed Next Steps for Adoption

If you agree with this analysis, we can immediately begin adopting the missing LLM WIKI concepts into DSOM by:

  1. Creating the palace-auditor skill: To allow me to "Lint" our knowledge base.
  2. Creating the dsom-knowledge-ingester skill: To handle the "Ingest" phase for external documents (like your NOSS integration).
  3. Adding a "Knowledge Compounding" rule to AGENTS.md: Instructing me to permanently save valuable chat analyses into the Palace.

Deep State of Mind (DSOM) For My AI Protocol | Harisfazillah Jamel (LinuxMalaysia) | 2026-07-16 Standard: UK English | DBP-standard Bahasa Melayu Malaysia (Piawai) | GNU General Public License v3.0