White Paper 1 of 3 — Vision & Thesis | MasteryMade / Forge Infrastructure | April 2026

The Intelligence Hub: Knowledge as a Compiler Problem

Why RAG is wrong, what Karpathy figured out, and how Forge goes further.

The Problem Everyone Solves Wrong

The entire AI industry has converged on RAG (Retrieval-Augmented Generation) as the answer to "how do I give an LLM my data?" The pattern: chunk your documents, embed them as vectors, retrieve relevant chunks at query time, paste them into the prompt.

It works. But it has a fundamental flaw: nothing compounds.

Every query starts from scratch. The LLM re-derives knowledge from raw documents each time. Cross-references aren't maintained. Contradictions aren't flagged. Synthesis doesn't accumulate. You're paying full cost for every question, and the system is exactly as smart on day 100 as day 1.

The Insight: Knowledge Is a Compiler Problem

Karpathy's reframe (April 2026): don't retrieve and generate. Compile.

Instead of re-deriving knowledge per query, the LLM incrementally builds and maintains a persistent wiki — a structured, interlinked collection of markdown files. When you add a new source, the LLM reads it, extracts key information, and integrates it into the existing wiki — updating entity pages, revising topic summaries, noting contradictions. The knowledge is synthesized once and kept current, not re-derived on every query.

Three Non-Negotiable Layers

1. Raw Sources (Immutable) — The source of truth. Append-only. The LLM reads from this but never modifies it. Session transcripts, intel signals, meeting notes, decision logs.
2. Compiled Knowledge (LLM-Maintained) — Structured articles: concepts, connections, Q&A, project syntheses. The LLM owns this layer. Humans read it; the LLM writes it.
3. Schema (Compiler Specification) — The rules that govern how compilation works. Article types, linking conventions, quality thresholds, conflict resolution.

Four Operations

1. INGEST — New source arrives → normalized to common format → stored in raw layer → compilation triggered
2. COMPILE — LLM reads schema + existing articles + new sources → creates/updates concept articles, connection articles, project syntheses → updates master index
3. QUERY — Ask anything → LLM reads index → drills into articles → synthesizes answer. The compounding loop: good answers get filed back as Q&A articles, making the system smarter with every question
4. LINT — Periodic health checks: broken links, orphan articles, contradictions, knowledge gaps. Self-correcting. Trustworthy.

Why Forge Goes Further

Karpathy's system is a personal tool for one researcher. Every implementation on the market (claude-memory-compiler, nvk/llm-wiki, second-brain, etc.) is single-user, single-session. None handle:

• Multiple agents writing to the same compiled knowledge base
• Cross-agent knowledge transfer (what Ralph learned is available to Claude Code sessions)
• Service-level query API (any microservice calls /v1/query)
• Graph intelligence (bridge detection, community clustering, gap analysis)
• Temporal reasoning ("what did we believe about X in March vs now?")

Personal Brain + Organizational Brain

Forge doesn't replace the personal second brain. It adds an organizational brain alongside it:

The KFS Advantage: Why We're Structurally Better

Karpathy's system is a flat wiki. Forge's Intelligence Hub is a wiki + entity graph + temporal graph + expert analysis. The Knowledge Fabric Service (KFS) is the storage and query engine that makes this possible.

The graph layer finds connections the compiler didn't explicitly create. It detects knowledge gaps — "you know about A and C but nothing about B which connects them." Temporal reasoning answers "what changed?" not just "what is?" These are impossible with text-only systems.

The Memex, Realized

Bush envisioned the Memex 81 years ago. The missing piece was always maintenance — who keeps the cross-references current, who flags contradictions, who writes the synthesis that connects disparate documents? Humans abandon knowledge bases because the maintenance burden grows faster than the value.

LLMs solve the maintenance problem. They don't get bored. They don't forget to update a cross-reference. They can touch 15 files in one pass. The wiki stays maintained because the cost of maintenance is near zero.

The human's job is to curate sources, direct the analysis, ask good questions, and think about what it all means. The LLM's job is everything else.

Forge adds the final piece Bush couldn't have imagined: multiple AI agents maintaining one shared Memex, each contributing from their domain. Not one researcher's brain — an entire organization's collective intelligence.

Published by Forge Intelligence Hub project · April 2026 · MasteryMade infrastructure

References: Karpathy LLM Wiki Gist (2026), Cole Medin claude-memory-compiler, Vannevar Bush "As We May Think" (1945)