Source-derived architecture field guide
Raw sources stay authoritative. Agents maintain the markdown memory layer. Routing files make the context traversable.
Most AI document workflows retrieve from raw source fragments at question time. An LLM wiki changes the operating model: it compiles sources into a persistent, cross-linked markdown knowledge layer that agents can maintain, humans can inspect, and future tools can traverse without starting over.
Source dossier
Reading posture
Use the video as a concrete workflow example. Use Karpathy's gist for the conceptual LLM Wiki frame. Use Obsidian documentation for vault, markdown, link, and graph mechanics.
Executive scan mode
This mode gives the thesis in under a minute: raw sources stay authoritative, wiki pages become maintained synthesis, and routing files turn memory into something agents can traverse instead of reconstructing.
What changed
The useful shift is not that a new model can make a prettier demo. The shift is that source material can be compiled into a maintained markdown knowledge layer instead of being rediscovered from raw fragments every time an agent answers a question.
Why it matters
A compiled wiki lets humans inspect pages, follow links, check the raw source boundary, and see how the knowledge base changed. That is materially better than invisible chat memory, opaque retrieval, or a pile of uploaded documents pretending to be architecture.
What to do next
Keep raw sources immutable. Let agents maintain derived wiki pages. Use routing files to define ingest, answer, update, log, and lint behavior. Review the synthesis before trusting it.
Thesis/article reading view
A coherent article path for readers who want the architecture without scanning diagrams first.
Uploading documents gives an agent access to material. It does not create durable knowledge. Without a maintained structure, each session has to rediscover what matters, infer relationships again, and rebuild a working model from scattered context.
That can be useful for one-off questions. It is weak as an operating model for accumulated knowledge. The failure mode is familiar: documents pile up, retrieval gets noisy, summaries drift, and nobody can tell which answer came from evidence versus whatever the model smoothed into confidence.
The better model separates three jobs. Raw sources are the evidence layer. They stay immutable and preserve provenance. The compiled wiki is the working memory layer. It contains derived markdown pages: summaries, source pages, entity pages, concept pages, comparisons, contradictions, and synthesis.
The routing schema is the operating contract. Files such as AGENTS.md, CLAUDE.md, index.md, and log.md tell the agent how to ingest sources, traverse the wiki, update pages, cite evidence, and record changes.
This is the control-plane distinction: source truth is not the same thing as agent memory.
The source video demonstrates one creator's workflow for turning transcripts, PDFs, URLs, and business context into an agent-maintained wiki. It is useful as a concrete applied example, not as a benchmark. It does not prove Fable is better than Opus, that every ingest will produce reliable pages, or that an AI OS becomes objectively smarter.
Karpathy's LLM Wiki gist gives the stronger architecture frame. The key distinction is persistent synthesis. Instead of treating raw sources as something to retrieve from at query time, the model incrementally maintains a structured, interlinked markdown wiki that compounds as sources are added and questions are answered.
That is the durable lesson: memory improves when it becomes a governed artifact, not when it becomes another hidden vendor feature.
Guided explainer mode
The LLM wiki is not a magic memory feature. It is a small file contract that makes context maintainable, inspectable, and traversable.
Anatomy
Larger systems can add receipts, source registries, contradiction queues, validators, and search. Those are governance controls, not decoration.
raw/Holds source truth. Raw evidence stays separate from the working memory layer.wiki/Holds derived memory: concepts, entities, sources, topics, comparisons, and synthesis pages.index.mdPoints agents and humans toward the right pages before they wander.log.mdRecords ingests, material updates, lint passes, and meaningful query work.AGENTS.md or CLAUDE.mdDefines allowed behavior, traversal order, citation posture, update rules, and escalation conditions.Flat versus structured
A flat wiki is better when the domain is small and speed matters. A structured wiki is better when recurring entities, concepts, systems, projects, decisions, and source types need different handling.
Use when the source set is limited, the domain is narrow, and taxonomy would make the tree look more impressive than useful.
Use when folders, page types, source registries, and lint checks reduce confusion instead of creating empty architecture.
Index, log, backlinks, traversal
A wiki is only useful to agents if the agent knows how to traverse it.
Evidence and implementation mode
Raw source is evidence. Wiki pages are derived memory. The public artifact keeps that boundary visible because LLM-written pages can be wrong.
It means recording source identity, preserving raw evidence, reviewing material synthesis, logging changes, and periodically linting for broken links, contradictions, stale claims, duplicate pages, and orphaned concepts.
Markdown can be inspected by humans, stored in Git, opened in editors, browsed in Obsidian, and read by different agents with file access. Different tools still handle files, links, permissions, and context limits differently.
Builder playbook
The durable idea is this: the model can change, but the memory layer should not reset.
Create an Obsidian vault or plain markdown folder. Add instructions, raw/, wiki/, index.md, and log.md. Add one source, ask the agent to ingest it, review the pages, tighten the schema, and repeat one source at a time.
Add a source registry, claim ledger, contradiction queue, deterministic validators, import/export rules, lint checks, receipts, and publication-safe boundaries if anything becomes public.
Risks and failure modes
The countermeasure is not more clever prompting. It is source discipline, page conventions, review gates, receipts, and validators.
| Failure mode | Control |
|---|---|
| Raw source loss | Preserve raw evidence separately from derived pages. |
| Private source leakage | Redact before ingest and keep public-safe boundaries explicit. |
| Taxonomy drift | Review page conventions and folder structure when confusion appears. |
| Stale claims | Use claim status, source dates, and review queues. |
| Broken links or orphan pages | Run link and orphan checks as boring maintenance. |
| Unsupported source-strength upgrades | Keep source claim, verified fact, teaching interpretation, and speculation separate. |
| UI polish masking weak evidence | Make evidence posture and provenance visible to readers. |
Evidence and provenance
The source video started the lane. The public artifact keeps the durable architecture pattern and leaves private preparation material out-of-band.
A public YouTube demonstration introduced the lane and supplied the concrete workflow: source material compiled into an agent-maintained markdown wiki. The raw source packet stayed out-of-band. ChatGPT normalized the automatic captions, extracted candidate claims, and separated source claims from teaching interpretation before publication.
Public sources verified the durable mechanics: Karpathy's LLM Wiki gist for the compiled-context idea and official Obsidian documentation for vaults, markdown files, internal links, and graph view. Model and product claims from the demonstration were softened. This page teaches the compiled-context pattern; it does not argue that Fable is superior, benchmark a model, or endorse a product.
The source video is useful because it shows one creator workflow for turning transcripts, documents, links, and business context into an agent-maintained wiki.
Karpathy's LLM Wiki gist supports the architecture frame: raw sources, persistent markdown wiki, schema, index, log, and maintenance loop.
Obsidian documentation supports the public mechanics: vaults, markdown files, internal links, and graph view.
Visual prompt suite
Each concept includes mainline white and dark expressive prompt variants. Prompt text is selectable and readable without JavaScript; copy buttons are progressive enhancement.
Raw sources, compiled markdown wiki, routing files, and agent traversal.
Render a mainline white 16:9 technical architecture diagram for engineering leaders and AI governance reviewers. Visible title: "Compiled Context Architecture". Visible subtitle: "Raw sources stay authoritative. The wiki becomes working memory. Routing files make traversal reliable." Layout: left-to-right three-layer system map with a raw source stack on the left, a cross-linked markdown wiki in the center, routing files above the path, an agent answer/update loop on the right, and a human review plus lint gate along the bottom. Labeled visual elements: "Raw source layer", "Compiled markdown wiki", "Routing and governance layer", "Agent traversal", "Immutable evidence", "Derived synthesis", "Index and log", "Human review", "Lint for drift". Palette and visual style: near-white background, charcoal text, quiet gray containers, red accent path, editorial technical field-guide style. Exact text constraints: render only the specified title, subtitle, labels, and callouts; keep all text horizontal and readable. Source/claim posture constraints: show the YouTube source only as applied demonstration context, not benchmark proof or product endorsement; do not claim Fable superiority or assured correctness. Do not depict fake screenshots, vendor dashboards, benchmark charts, secret source files, local paths, transcript snippets, or raw source packets. Accessibility / alt-text concept: three-layer LLM wiki architecture showing raw evidence, compiled markdown memory, and routing governance. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".Render a dark expressive 16:9 technical architecture diagram for builders evaluating agent memory systems. Visible title: "Compiled Context Architecture". Visible subtitle: "The model can change. The markdown memory layer persists." Layout: graphite architecture board with luminous raw evidence files on the left, a stable compiled wiki graph in the center, routing schema cards above, an agent loop on the right, and a red governed path through index, answer, update, and review. Labeled visual elements: "Raw evidence", "Compiled wiki", "Routing schema", "Agent loop", "Review gate", "Do not rewrite source truth", "Maintain synthesis", "Traverse by index", "Append to log", "Escalate contradictions". Palette and visual style: black, graphite, near-white, muted gray, red accents, high-contrast technical poster. Exact text constraints: render only the specified title, subtitle, labels, and callouts; no decorative microtext. Source/claim posture constraints: represent compiled context as an architecture pattern, not as a model benchmark, Fable endorsement, or proof of superior output. Do not depict product UI, fake benchmark scoreboards, raw transcripts, source packet archives, local file paths, or proprietary material. Accessibility / alt-text concept: dark architecture board showing portable compiled context controlled by routing and review. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".Disposable query-time reconstruction versus maintained synthesis.
Render a mainline white 16:9 comparison diagram for technical decision makers choosing between query-time retrieval and maintained memory. Visible title: "Retrieval-Only vs LLM Wiki". Visible subtitle: "One reconstructs context at query time. The other maintains compiled synthesis." Layout: two balanced columns separated by a red boundary line; the left column shows a fragmented loop that resets, and the right column shows a maintained wiki layer that accumulates reviewable context. Labeled visual elements: left header "Retrieval-only loop"; right header "LLM wiki loop"; left labels "Query", "Chunk search", "Temporary answer", "Context resets"; right labels "Ingest source", "Update wiki", "Update index and log", "Answer from maintained context"; callouts "Repeated reconstruction", "Persistent synthesis", "Reviewable memory". Palette and visual style: near-white background, charcoal text, gray system boxes, red boundary and control marks. Exact text constraints: render only the specified title, subtitle, headers, labels, and callouts. Source/claim posture constraints: compare operating models only; do not imply universal accuracy, speed, safety, or business outcome gains. Do not depict fake screenshots, model leaderboards, product logos, source packets, raw transcript text, or local paths. Accessibility / alt-text concept: comparison of disposable retrieval and persistent LLM wiki memory. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".Render a dark expressive 16:9 comparison diagram for builders explaining why compiled context differs from repeated retrieval. Visible title: "Retrieval-Only vs LLM Wiki". Visible subtitle: "Disposable context on the left. Compiled memory on the right." Layout: left side uses dissolving fragments and a short reset loop; right side uses a stable markdown graph with cross-links, index, log, and review rail. Labeled visual elements: left header "Retrieve again"; right header "Maintain and traverse"; labels "Raw chunks", "Temporary synthesis", "Persistent pages", "Cross-links", "Index", "Log", "Review". Palette and visual style: black and graphite background, near-white labels, muted gray containers, red route line. Exact text constraints: render only the specified title, subtitle, headers, and labels. Source/claim posture constraints: present this as an architecture tradeoff, not a claim that one vendor, model, or product always performs better. Do not depict fake benchmark charts, vendor logos, product screenshots, raw transcripts, source packet files, local paths, or private data. Accessibility / alt-text concept: dark comparison between repeated retrieval and maintained wiki context. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".The small file contract that turns memory into something agents can traverse.
Render a mainline white 16:9 folder anatomy diagram for teams designing a minimal agent-readable wiki. Visible title: "Anatomy of an LLM Wiki". Visible subtitle: "A small file contract turns memory into something agents can traverse." Layout: clean file-tree diagram on the left, callout rail on the right, and one red traversal arrow that starts at the routing file, passes through index and log, and reaches derived wiki pages. Labeled visual elements: "raw/", "wiki/", "concepts/", "entities/", "sources/", "index.md", "log.md", "AGENTS.md or CLAUDE.md", "Source truth", "Derived pages", "Navigation", "Chronology", "Operating contract". Palette and visual style: near-white, charcoal, gray, red accent, precise editorial schematic. Exact text constraints: render only the specified title, subtitle, folder labels, and callouts. Source/claim posture constraints: show a teaching structure, not a required standard or guarantee that agents will update correctly. Do not depict actual local directories, raw source packet names, transcript excerpts, hidden files, fake terminal screenshots, or product-specific UI. Accessibility / alt-text concept: file tree showing the minimum structure of an LLM wiki. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".Render a dark expressive 16:9 file-system architecture diagram for architects explaining the control-plane role of plain files. Visible title: "Anatomy of an LLM Wiki". Visible subtitle: "The boring files are the control plane." Layout: dark terminal-like file tree with a red routing line, subtle graph nodes around derived wiki pages, and a compact governance strip under the tree. Labeled visual elements: "raw/", "wiki/", "index.md", "log.md", "AGENTS.md", "source registry", "lint checks", "Immutable", "Derived", "Navigable", "Auditable", "Reviewable". Palette and visual style: graphite, black, near-white, gray, red control accents. Exact text constraints: render only the specified title, subtitle, labels, and callouts. Source/claim posture constraints: represent file conventions as a practical pattern, not an official Obsidian requirement, model benchmark, or product claim. Do not depict real private paths, source packet archives, raw transcript text, fake command output, vendor dashboards, or unsupported model names. Accessibility / alt-text concept: dark file tree showing LLM wiki source, memory, and governance files. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".A governed agent reads the route before it answers.
Render a mainline white 16:9 process diagram for AI governance reviewers defining how an agent should traverse maintained context. Visible title: "Agent Traversal Path". Visible subtitle: "A governed agent reads the route before it answers." Layout: left-to-right process rail with eight numbered steps, red control checkpoints at source check, log check, allowed update, and receipt, plus three callout flags above the rail. Labeled visual elements: "1 Read routing file", "2 Read index", "3 Select pages", "4 Check sources", "5 Check log", "6 Answer with posture", "7 Update only if allowed", "8 Append receipt", "Do not wander blindly", "Escalate contradictions", "Preserve source truth". Palette and visual style: near-white, charcoal, gray, red checkpoint accents, operations-manual clarity. Exact text constraints: render only the specified title, subtitle, steps, and callouts. Source/claim posture constraints: show a governed workflow recommendation, not a guarantee that agents will comply or that all answers become accurate. Do not depict fake chat transcripts, private source snippets, raw transcript text, source packet names, local paths, or autonomous high-risk actions. Accessibility / alt-text concept: sequential traversal contract for an agent using an LLM wiki. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".Render a dark expressive 16:9 process diagram for builders presenting a governed agent route through wiki memory. Visible title: "Agent Traversal Path". Visible subtitle: "Memory is only useful when the route is explicit." Layout: dark architecture rail with luminous nodes, a red guardrail path, and three review checkpoints that interrupt the path before answer, update, and receipt. Labeled visual elements: "Routing", "Index", "Candidate pages", "Source check", "Log check", "Answer", "Allowed update", "Receipt", "Bounded traversal", "Claim posture", "Human review". Palette and visual style: black, graphite, near-white, gray, red guardrail path, cinematic but readable. Exact text constraints: render only the specified title, subtitle, labels, and callouts. Source/claim posture constraints: frame agent traversal as controlled process design, not autonomous authority, benchmark evidence, or product endorsement. Do not depict fake product UI, benchmark trophies, raw source files, transcripts, local paths, source packet archives, or unreviewed private data. Accessibility / alt-text concept: dark process path showing governed agent navigation through wiki memory. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".Risks and controls for a wiki that compounds over time.
Render a mainline white 16:9 governance risk matrix for reviewers checking whether compiled context has enough controls. Visible title: "LLM Wiki Failure Modes". Visible subtitle: "Compiled context needs controls, not vibes." Layout: two-column risk/control matrix with eight rows, red risk markers on the left, gray control rails on the right, and a thin source-boundary line across the bottom. Labeled visual elements: risk labels "Raw source loss", "Private data leakage", "Taxonomy drift", "Stale claims", "Broken links", "Orphan pages", "Unsupported claims", "Unreviewed batch ingest"; control labels "Preserve raw", "Redact before ingest", "Schema review", "Claim status", "Link lint", "Orphan check", "Evidence ledger", "Human gate". Palette and visual style: near-white, charcoal, gray, red, compact governance dashboard without fake product chrome. Exact text constraints: render only the specified title, subtitle, risk labels, and control labels. Source/claim posture constraints: treat risks as governance concerns, not proof that any specific product is unsafe or superior. Do not depict real user data, raw transcripts, private source material, source packet names, local paths, vendor logos, or model benchmark rankings. Accessibility / alt-text concept: matrix pairing LLM wiki risks with governance controls. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".Render a dark expressive 16:9 governance risk board for teams teaching why maintained wiki memory needs review gates. Visible title: "LLM Wiki Failure Modes". Visible subtitle: "The wiki compounds. So do the mistakes." Layout: dark risk board with six risk cards on the upper row, six matching control blocks on the lower row, red warning rails connecting each pair, and a source-boundary strip at the bottom. Labeled visual elements: risk labels "Drift", "Stale claims", "Hallucinated links", "Raw source loss", "Private leakage", "UI polish over evidence"; control labels "Review", "Lint", "Receipts", "Evidence ledger", "Redaction", "Source boundary". Palette and visual style: black, graphite, near-white, gray, red warning accents, high-contrast operations board. Exact text constraints: render only the specified title, subtitle, risk labels, and control labels. Source/claim posture constraints: explain control needs without claiming benchmark results, product superiority, or assured safe behavior. Do not depict private data, raw transcript excerpts, local file paths, source packet archives, fake screenshots, or vendor logos. Accessibility / alt-text concept: dark governance board showing risks and controls for compiled wiki memory. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".The full artifact summarized as one architecture poster.
Render a mainline white 16:9 capstone architecture poster for executives and builders summarizing the full SharePlane lesson. Visible title: "The Model Can Change. The Memory Layer Should Not Reset." Visible subtitle: "Own the compiled context. Rent the intelligence." Layout: stable memory stack in the center, raw source layer at the base, markdown wiki and routing files in the middle, review gates across the front, and multiple interchangeable agent/model slots connected by a red portability path. Labeled visual elements: "Raw sources", "Markdown wiki", "Routing files", "Review gates", "Swappable agents", "Same evidence", "Same routes", "Same permissions", "Different model". Palette and visual style: near-white, charcoal, gray, red portability path, premium editorial poster with precise labels. Exact text constraints: render only the specified title, subtitle, labels, and callouts. Source/claim posture constraints: present a governance and portability pattern, not a benchmark, product endorsement, or claim that any model always improves with this setup. Do not depict fake model scorecards, vendor dashboards, raw transcripts, source packet names, local paths, screenshots, or secret documents. Accessibility / alt-text concept: portable markdown memory layer connecting to different agent models without losing context. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".Render a dark expressive 16:9 capstone architecture poster for technical leaders explaining why owned context should outlast model changes. Visible title: "The Model Can Change. The Memory Layer Should Not Reset." Visible subtitle: "Own the compiled context. Rent the intelligence." Layout: dark stable memory stack in the center, red routing path around it, replaceable model slots on the right, human review gate in front, and source-truth foundation under the stack. Labeled visual elements: "Raw evidence", "Markdown memory", "Routing schema", "Review gates", "Swappable agents", "Same source truth", "Derived memory", "Index and log", "Different model". Palette and visual style: black, graphite, near-white, gray, red control path, premium dark technical poster with readable labels and strong hierarchy. Exact text constraints: render only the specified title, subtitle, labels, and callouts. Source/claim posture constraints: show stable compiled context as a reusable architecture pattern; do not claim Fable superiority, model benchmark wins, assured safe behavior, or product endorsement. Do not depict fake screenshots, benchmark charts, vendor logos, raw transcript text, source packet archives, local paths, or private documents. Accessibility / alt-text concept: dark poster showing stable compiled context outlasting model changes. Do not render meta labels such as "prompt", "metadata", "layout notes", "aspect ratio", "alt text", or "visual prompt".