Compiled Knowledge Systems

A compiled knowledge system turns raw capture into durable understanding by maintaining a smaller, linked synthesis layer that improves through ingest, curation, query, and file-back.

Visual navigation Use the cluster tools to review this hub as a navigable system, not only as prose: - Compiled Knowledge Cluster Dashboard - Compiled Knowledge Cluster - Local visuals

!Illustrative compiled knowledge loop

*Generated conceptual illustration, not source evidence: raw captures become a curated synthesis map, then loop back through dashboards, validation, and file-back.*

flowchart TD
    A["Raw evidence"] --> B["Ingest"]
    B --> C["Compiled wiki"]
    C --> D["Hub and concept pages"]
    D --> E["Query from synthesis first"]
    E --> F["Useful new synthesis"]
    F --> G["File back into wiki"]
    G --> C
    C --> H["Lint and curation"]
    H --> I["Merge, demote, archive, relink"]
    I --> C

Why this matters

The vault should not behave like a RAG bucket, a clipping dump, or a graph visualization demo. Its value comes from compounding editorial synthesis.

The core move is simple: instead of making the model rediscover the same ideas from raw sources every time, the system writes durable pages that preserve the best conceptual work. Later queries start from the maintained wiki, then fall back to raw evidence when needed.

The newest Codex workflow sources add an operational version of the same pattern: messy data and scattered feedback should be turned into reviewable artifacts rather than overwritten or summarized away. A good compiled knowledge system preserves raw evidence, creates a cleaned or synthesized copy, records confidence and source links, and leaves the next action visible.

A newer Codex skills source adds the maintenance loop this vault now uses: when a compile workflow works, turn the thread, commands, scripts, and quality bar into a skill so future ingest is consistent. A newer world-knowledge paper adds an adjacent research pattern: agents can explore a site or environment and compile a guidebook before later tasks need it.

A newer Karpathy interview source reinforces the original reason this hub exists. A model can recompile documents into a wiki that becomes easier for both humans and later agents to read. The point is not to replace understanding with storage; it is to maintain a persistent synthesis layer so each later question starts from accumulated structure instead of raw rediscovery.

This page organizes:

• LLM Memory
• Second Brain Systems
• Personal Knowledge Systems
• AI-Assisted Content Systems
• Chief of Staff Agents
• Agent Memory & Context Systems

Core thesis

Knowledge systems become useful when they separate evidence from synthesis.

The wiki should improve human understanding, not merely increase recall. Source ingestion is valuable when it changes the durable map: a hub gets sharper, a concept gets better boundaries, a contradiction is recorded, or a useful answer becomes a file-backed artifact.

Implementation model

1. Ingest should update the map, not just add pages

New sources should usually improve existing pages. A new first-class page is justified only when the source reveals a durable concept, workflow, or case that cannot be handled inside the current map.

The raw/ folder is intentionally an inbox plus evidence layer. Sources should remain there after successful compile; the manifest, not file movement, determines whether a source is new, changed, already represented, compiled, quarantined, or deferred.

That creates a simple daily rhythm:

flowchart TD
  A["Drop source into raw"] --> B["Audit manifest status"]
  B --> C{"New or changed?"}
  C -->|No| D["Leave source as evidence"]
  C -->|Yes| E["Compile into wiki synthesis"]
  E --> F["Update source notes and manifest"]
  F --> G["Lint links, raw refs, assets, and orphans"]
  G --> D

This preserves Karpathy's simple source/wiki split while still supporting incremental scheduled ingest.

2. Hubs should curate

A hub is not a folder index. It should:

• state the domain thesis
• show a visual map
• name the best reading path
• explain why adjacent pages matter
• route narrow details through supporting-page sections in the index

3. Concepts should earn their place

A concept page should answer:

What does this help the reader understand that raw sources alone do not make obvious?

If a page mostly preserves a single source, it belongs as reference or case material.

4. Retrieval is not enough

Retrieval is valuable for recall, precision, and evidence. But the compiled layer adds something retrieval alone does not:

• reconciled claims
• cross-source synthesis
• durable terminology
• reading paths
• editorial judgment
• remembered conclusions

5. The system should learn from use

Strong knowledge systems have learning loops:

• ingest new sources
• answer questions from the wiki first
• file back strong answers
• lint weak links and stale claims
• merge or demote pages when structure gets flat
• promote repeatable maintenance workflows into skills

6. Feedback synthesis is a compile operation

The feedback-synthesis source is a practical example of compilation outside a wiki:

1. gather feedback from Slack, issues, surveys, notes, or exports
2. cluster repeated themes
3. retain source links or IDs
4. mark confidence and ambiguity
5. separate product follow-up from engineering follow-up
6. draft the next artifact only after review

That is the same pattern this vault should use for raw sources. The output should not be a prettier inbox. It should be a reviewable synthesis layer with provenance.

7. Cleaning data is preservation plus transformation

The messy-data source adds a useful rule for all compiled systems:

Keep the original unchanged; write the improved copy separately.

For this vault, that means:

• raw/ stays immutable evidence
• wiki/ carries enriched synthesis
• reports explain what changed
• uncertain or low-confidence material stays visible instead of silently disappearing

The spreadsheet-cleaning example makes that rule more concrete. A good derivative artifact should preserve source row IDs when possible, normalize only the requested fields, and emit a short data-quality note that distinguishes:

• rows changed confidently
• rows removed as duplicates or summary noise
• rows that could not be normalized cleanly

That distinction matters because transformation without an uncertainty trail turns maintenance into silent corruption.

8. Successful workflows should compile into skills

The reusable-skills source adds a meta-rule for this vault:

If an agent workflow becomes useful and repeatable, compile the workflow itself.

That means preserving:

• the trigger
• the source material the agent should read
• scripts and commands that should be reused
• examples of good outputs
• validation checks
• failure corrections discovered in later runs

This is why /kb-compile belongs as a skill-backed command rather than as a loose prompt. It turns maintenance from heroic one-off editing into an improving operating procedure.

9. Environment guidebooks are compiled knowledge artifacts

The world-knowledge paper describes a pattern close to a domain-specific mini-wiki: an agent explores a website, groups pages, summarizes useful facts, and produces a guidebook that later tasks can use.

The useful rule for this vault is:

• raw browsing or crawling is evidence
• the guidebook is compiled knowledge
• later task execution should start from the guidebook, then fall back to fresh inspection when the guidebook is insufficient

That same pattern applies to codebases, websites, Obsidian vaults, vendor docs, and local workflows.

Visual and editorial standard

Use visuals when they explain structure:

• Mermaid maps for systems and workflows
• tables for distinctions and decision rules
• screenshots for real interfaces or evidence artifacts
• source images for infrastructure, mission systems, or product surfaces
• generated conceptual images for teaching aids and memory hooks, with clear "not evidence" provenance

Avoid decorative imagery. A beautiful knowledge base is one where structure, priority, and evidence are immediately legible.

Visual reading cues

Read the image as the mental model for this whole page:

Failure modes

• One page per source.
• A flat index that lists everything.
• Link density without a reading path.
• Graph aesthetics without editorial judgment.
• Ingest that only appends and never consolidates.
• Treating source notes as enough when the page lacks insight.
• Forgetting that raw/ is evidence and wiki/ is synthesis.
• Overwriting or hiding the messy source instead of preserving it beside the cleaned synthesis.
• Leaving successful maintenance workflows trapped in chat instead of compiling them into skills.
• Treating an environment guidebook as ground truth after the environment changes.
• Cleaning operational data in place with no derivative artifact or uncertainty note.

Read next

1. Agent Memory & Context Systems for memory, retrieval, compaction, and persistence.
2. Personal Knowledge Systems for local vault and domain-specific system design.
3. Second Brain Systems for operator support across long work histories.
4. AI-Assisted Content Systems for publishing and connection-finding workflows.
5. Agent Execution Systems for turning maintenance loops into repeatable agent workflows.