Knowledge Compilation
Knowledge compilation is the process of turning raw sources and session evidence into durable, linked, current wiki truth that future agents can retrieve instead of re-deriving.
Core model
Raw data is not memory. A source becomes memory only after it is filed, synthesized, linked, and made retrievable.
This is the wiki's advantage over ordinary RAG. RAG retrieves source chunks and asks the model to synthesize live. The wiki compiles the synthesis once, keeps evidence in timelines, and makes the current answer retrievable.
Compiled truth vs timeline
| Layer | Mutability | Purpose |
|---|---|---|
| Compiled truth | Rewrite when evidence changes | Current best understanding, organized by theme |
| Timeline | Append-only | Dated evidence trail and source provenance |
| Raw source | Immutable | Original material, not edited by agents |
| Index/search | Regenerated | Retrieval surface for future sessions |
The failure mode is append-only knowledge. Appending every new source to the bottom of a page leaves future agents with archaeology, not memory.
Section implications
| Section | Compilation behavior |
|---|---|
concepts/ |
Extract reusable primitives from repeated project/tool/design evidence. |
projects/ |
Rewrite current status and product thesis; append launch/metric/decision evidence. |
tools/ |
Keep current capability, routing, and tradeoff facts verified from primary docs. |
skills/ + Skill Registry |
Keep executable procedures in skills/**/SKILL.md; expose inventory through generated registry rows and compact family synthesis pages. |
design/ |
Convert inspiration into constraints, tokens, and verification rules. |
research/ |
Preserve paper/source facts, then promote reusable lessons into concept pages. |
meta/ |
Document the operation, not the general philosophy unless it changes agent behavior. |
The compilation checklist
- Search existing pages first.
- Decide whether the source changes compiled truth or only adds timeline evidence.
- Update all primary and related pages, not only the obvious one.
- Add wikilinks where the concept should be discoverable.
- Keep claims source-backed.
- Run
build-index, doctor,qmd update, andqmd embed. - Log the operation when it changes the wiki's durable state.
Why it matters for agents
Agents are cheap at synthesis once. They are expensive at repeated synthesis forever. Knowledge compilation pays the cost once and amortizes it across every future query, skill run, design pass, project briefing, and career narrative.
Kevin should not have to remember which source said what. The wiki should answer from compiled truth and expose the timeline when proof matters.
Concept Position
| Field | Value |
|---|---|
| Concept family | Brain, memory, and retrieval |
| Concept owned | Knowledge compilation is the process of turning raw sources and session evidence into durable, linked, current wiki truth that future agent... |
| Category map | Concept System Map |
Company-scale compilation (Ramp Glass)
Ramp's Glass applies knowledge compilation to a company. The launch essay's premise: giving every employee the best models is not the same as becoming AI-native, because each person's best prompts and workflows live in one chat history and vanish. Glass makes the company's learned answers the starting point instead of an empty chat. Finance taught it once to exclude U.S. companies with U.K. offices from U.K. customer counts, and everyone's later questions get that rule. 323 employees made over 6,500 edits to what it knows. One saved prompt took feature-request capture from about 300 a month by hand to 20,000 pulled from customer calls. Source: X/@eglyman, 2026-10-13; Source: Ramp Glass
The rule matches this wiki's: learn a thing once, write it where the next run starts. Learn From Your Own Record collects prompts that mine one person's own session history the same way.
WikiSkill
Google's WikiSkill paper separates three things skill-evolution systems usually merge: raw execution traces, a persistent wiki of accumulated knowledge, and the executable skills. Experience is consolidated into the wiki, and each skill update builds on the wiki. Ablations show the wiki carries much of the gain; small models with evolved skills beat larger ones without them, and evolved skills transfer across model families. This is the structure of this wiki and its skills/ tree. Source: X/@dair_ai, 2026-08-28; Source: arXiv 2608.27454
A one-prompt "life OS" for a personal agent builds four linked layers: a Markdown knowledge base as source of truth, a persistent pointer so every session finds it, skills to capture, retrieve and correct context, and a private tracker for habits, goals, projects and reviews. Its best instruction is to inspect first and extend any existing knowledge base rather than create a competing one. Source: X/@shannholmberg, 2026-08-10
Cerebras's internal knowledge base drew 15,000 questions a day from employees, automations and agents within three months. Slite's Christophe Pasquier argues most teams should buy such a system rather than build it. Source: X/@Christophepas, 2026-07-22
A business version: keep offers, clients, voice, playbooks and "rulings" in one vault; give each worker only its slice; and turn every correction into one line in the rulings note that every worker reads first, so a mistake is fixed once. The workers are interchangeable; the vault is the business. Source: X/@EXM7777, 2026-07-25
The minimum context for a business workspace is four files: business.md, customer.md, offer.md and voice.md, including what the business must never sound like. Source: X/@coreyganim, 2026-07-20
The Cerebras guide covers the hard part: ingesting from Slack, code, docs, Jira and everything else, reconciling sources, ordering them with the right precision and recall, and keeping context fresh for agents. Source: X/@LinghuaJ, 2026-07-16
How the Cerebras base works: leave data where it lives (Slack, GitHub, Jira, docs) instead of centralizing; for Slack, refetch whole threads, extract question, summary and resolution, and embed bursts of messages; search with full text, embeddings, IDF and age decay together; index code incrementally; and answer with a planner that fans out to parallel tools, merges with reciprocal rank fusion and a reranker, then synthesizes. No knowledge graph, and agents call the same tools over MCP. Source: X/@Krongggggg, 2026-07-16
Three layers often confused: an ontology is the rule book of what kinds of things exist and how they may connect, with no data; a knowledge base is the content, structured or not; a company brain is the system people and agents talk to, usually a model plus retrieval, tools and memory. The best systems will increasingly combine all three. Source: X/@KSimback, 2026-07-17 Microsoft open-sourced Ontology Playground, an ontology tool one post billed as a Palantir rival. Source: X/@IntuitMachine, 2026-07-18
Two lessons from the Cerebras write-up that are easy to miss: code search kept ripgrep next to embeddings (both, not either), and freshness decides survival, so only changed chunks are re-embedded and sync state lives with the vectors. Production retrieval is mostly tuning chunking, sync, reranking, tool APIs and permissions. Source: X/@sdhilip, 2026-07-17; Source: X/@jerryjliu0, 2026-07-18
Timeline
- 2026-07-01 | Concepts category refresh added this page to the Brain, memory, and retrieval family, linked it to Concept System Map, and kept it standalone because it owns this reusable mental model: Knowledge compilation is the process of turning raw sources and session evidence into durable, linked, current wiki truth that future agent... private page
- 2026-06-19 | Created after a whole-wiki audit found "compiled truth" referenced across meta/workflow/memory pages without a canonical concept page. Source: whole-wiki concept review, 2026-06-19