Scaling and Enterprise Systems
Articles exploring scaling limits, context-window optimizations, concurrency, access governance, and enterprise architectures.

An architectural analysis of how four builders (Andrej Karpathy, Garry Tan, Hannah Stulberg, and Ramp) converge on a shared organizational AI memory architecture. Identifies five universal pillars—a structured store, a routing layer, skills as the unit of work, a writeback habit, and separating context from compute—while analyzing practical deployment hurdles, capture friction, and permission boundaries.
Company Brain for Enterprise AI: Why the Data Layer Decides Everything (Mayank Mudgal on Colrows)colrows.comAn architectural essay examining why standard document wikis, vector RAG, and MCP tool registries fail as enterprise company brains. Argues that organizational AI agents require meaning to be a compile-time concern, proposing a governed semantic layer of typed entities, versioned metric definitions, and verifiable policy constraints to eliminate query-time hallucinations and metric disagreements.
Company Brain, Part 2: Factual Memory (Ashwin Gopinath)nanothoughts.substack.comAn architectural essay examining factual memory as the foundational layer of enterprise company brains. Argues against centralized repositories and naive RAG in favor of a semantic file system that maps artifact relationships, supports emergent memory from individual to team levels, and personalizes proactive context across IC, manager, and CEO roles.
Company Brain, Part 3: Interaction Memory (Ashwin Gopinath)nanothoughts.substack.comAn architectural essay exploring interaction memory as organizational chain-of-thought. Examines how to capture unwritten rationale, fragile assumptions, and implied commitments from human conversations into an evolving context graph, using domain ontologies to surface conflicting team assumptions and prevent context loss for autonomous agents.
Company Brain, Part 4: Action Memory (Ashwin Gopinath)nanothoughts.substack.comAn architectural essay exploring action memory as the operational layer of enterprise company brains. Deconstructs action into four components—procedural, trigger, execution, and outcome memory—arguing that 'doing nothing is a first-class action' and outlining contextual guardrails that govern when autonomous agents should act, wait, or escalate.

An architectural guide breaking down five enterprise company brain architectures: enterprise search, document RAG, memory APIs, live MCP tool connectors, and agent-written Markdown wikis. Evaluates synthesis timing, token economics, multi-tenant failure modes, and schema design for shared organizational vaults.
Company Brain: Why Most Companies Have Data But No Memory (Ashwin Gopinath)nanothoughts.substack.comAn architectural essay exploring the structural requirements of enterprise company brains. Formulates a four-layer framework (factual memory, human communication, context graphs, and action coordination) to distinguish passive search from living organizational memory that preserves decision rationale, tradeoffs, and intent for autonomous agents.
Designing a Medium, Not a Tool (Gurden Batra on Civilizational Optionality)civilizationaloptionality.substack.comAn architectural essay reframing the LLM Wiki as institutional learning infrastructure. Formulates three layers with distinct mutation rights (
raw/,wiki/, and policy), separates tripartite authority (deterministic instruments, model interpretation, human Git authorization), details deployment gates for public/private visibility tiers, and explores lexical centre-of-mass metrics to detect unresisted conceptual drift.
An architectural essay and demo comparing personal Obsidian vaults to enterprise company brains. Details how to adapt Karpathy's LLM Wiki pattern to multi-agent organizations using typed graph schemas as data contracts, open columnar object storage (Lance/DataFusion), and Git-style branch-and-merge workflows that enable concurrent agent writes while blocking conflicting updates for human review.

An architectural guide examining the transition from personal LLM wikis to scheduled, multi-agent company brains. Details access governance via scoped directory permissions, read-only external mounts (Google Drive, Notion), background cron maintenance loops, and agent-generated adaptive HTML interfaces using Cabinet.
Karpathy's LLM Wiki v2: What to Keeptheaioperator.ioA critical architectural teardown outlining the evolution of the LLM Wiki design pattern. Details the necessary shift from manual workflows to parallel background agent pipelines, explains how to replace heavy index rewrites with vector mapping arrays, and demonstrates how to implement a targeted JSON pre-routing layer to keep API context costs minimal as local vaults grow past thousands of nodes.

An architectural essay analyzing the three layers and operations of Karpathy's LLM Wiki pattern, examining empirical studies on token trade-offs and index navigation limits, and detailing the governance primitives (ingestion gates, claim expiration dates, and served-versus-cited metrics) required when scaling to shared multi-agent team wikis.
LLM Wiki v2: What Breaks at Scale (Rohit Ghumare)aiengineeringfromscratch.comAn architectural guide examining how flat markdown wikis degrade across thousands of sessions. Details memory lifecycle mechanics (confidence decay, Ebbinghaus retention curves, supersession), multi-tier consolidation, typed knowledge graphs, event-driven hooks, and hybrid search (BM25, dense vector, and graph traversal fused via RRF).
LLM Wiki: Karpathy's Pattern, Built for a Paying Client (Adam Sandler on The Viable Edge)viableedge.comAn architectural guide exploring how to adapt Karpathy's LLM Wiki pattern and Claude Code for client delivery. Details four key shifts for enterprise company brains: resolving conflicting client sources with provenance precedence rules, treating the schema as a contract, converting linting into structured retrieval test scorecards, and establishing autonomous handoff routines.

A conceptual primer analyzing the scaling of the LLM Wiki pattern to enterprise teams, detailing the compute cost tradeoffs (compile-time vs. query-time RAG), scheduled linting loops, access governance, version control integration, and agent memory architectures.
Meaning, Traversal, Reasoning, Action: The Graph Stack for Agentic AI (George Anadiotis in The Year of the Graph)yearofthegraph.xyzAn architectural analysis examining the convergence of flat-file LLM wikis, formal ontologies, and context graphs. Explores Vault-LD (Markdown with W3C YAML-LD frontmatter), Google's OKF v0.2 trust signals, bitemporal agent memory properties, and syncing Neo4j graphs deterministically from markdown directories.

An architectural essay examining the requirements for adapting Karpathy's personal LLM knowledge base pattern to enterprise organizations, detailing multi-source automated ingestion, factual verification layers, and self-healing staleness detection.

An architectural deep-dive exploring how to overcome the file-system and context-window scaling bottlenecks of Karpathy's LLM Wiki pattern using a lightweight property graph. Details how modeling containment, section sequences, and wikilinks in Neo4j enables constant-time progressive retrieval, graph traversal, and GDS community detection across large Markdown vaults without index-time token costs.

An architectural essay analyzing how to scale Karpathy's LLM Wiki pattern into organizational context infrastructure. Examines content entropy, drift, and ownership through a ContentOps lens, detailing requirements for non-technical approval workflows, audit trails, pinned overrides that survive rebuilds, and multi-audience wiki compilation.

An architectural guide unifying personal second brains, company brains, and multi-agent shared brains into a cohesive three-tier stack. Analyzes how a shared memory primitive applies across individual, team, and agent fleet scopes, explaining why naive RAG fails and detailing auto-consolidating memory, contradiction reconciliation, and multi-scope tagging.
The LLM Wiki at Scale: From Personal Research Tool to Production RAG (Michal Nasternak on Medium)michalnasternak.medium.comAn architectural analysis examining how to scale the LLM Wiki pattern to enterprise production. Details the bottlenecks of single index files once vaults grow past hundreds of pages, proposing session-level on-demand wiki compilation and per-user persistent knowledge layers to optimize synthesis quality, cost efficiency, and debuggability over traditional vector RAG.

An architectural primer proposing a dual-layer structure that pairs PostgreSQL (for transactional data/metrics) with a Markdown wiki (for qualitative context), detailing saved SQL queries, database branching, and agent skill configurations.

An architectural guide by Christophe Pasquier tracing the evolution of company brains from personal second brains, enterprise search, and Karpathy's LLM wiki pattern to multi-agent organizational context layers. Contrasts personal memory vaults against company brains across access permissions, conflicting source resolution, and automated maintenance, providing an operational decision framework for teams deploying autonomous agents.

An architectural guide defining the company brain category across five requirements: write-time comprehension, bi-temporal validity tracking, org-wide scope, commitment tracking, and proactive contradiction detection. Deconstructs four CoALA-derived memory layers (factual, interaction, reasoning, action) and outlines an implementation sequence to unify human and multi-agent context over MCP.
What Is an LLM Wiki and How to Build One for Your Team (Roop Reddy on Documentation.AI)documentation.aiAn architectural guide examining how to scale Karpathy's LLM Wiki pattern from individual vaults to collaborative team environments. Details managing context-window saturation past the 100-page threshold using integrated page search, establishing Git branch-and-merge review workflows for agent-drafted diffs, splitting human vs. model page ownership, and serving verified context to agents via Authoring and Reader MCP servers.

A deep architectural critique outlining solutions for token scaling limits in file-based context stores, featuring code patterns for semantic deduplication and pre-commit syntax hooks to protect structural integrity.
