Agent-Maintained Wiki

Source Excerpt

Category: Agent Pattern / Knowledge Management

Maturity: Established

Executive Summary

An agent-maintained wiki is a knowledge base where an Large Language Model agent autonomously writes, updates, and maintains all content—summarizing sources, creating entity/concept pages, cross-referencing, and health-checking for contradictions. Humans curate sources and ask questions; the agent does all bookkeeping. This pattern makes persistent wikis viable by eliminating the maintenance burden that causes human abandonment.

Definition

An agent-maintained wiki is a wiki where:

• The Large Language Model agent owns the entire wiki layer—creating pages, updating content, maintaining consistency
• Humans never (or rarely) write wiki content directly—they curate sources and ask questions
• The agent follows a schema document that specifies structure, conventions, and workflows
• Maintenance operations (cross-referencing, updating summaries, flagging contradictions) are automated

This is distinct from traditional wikis where humans write and maintain content, or RAG systems where no wiki exists at all.

Context & Motivation

The Problem

Traditional wikis fail because:

• Maintenance Burden: Updating cross-references, keeping summaries current, noting contradictions across pages—this tedious work grows faster than the wiki's value
• Human Abandonment: People prefer creating new content over maintaining old content. Wikis become stale and inconsistent.
• Inconsistent Quality: Different authors, different styles, varying levels of detail and accuracy

The Solution

Large Language Model agents solve these problems:

• No Boredom: Agents don't get bored updating cross-references or filing summaries consistently
• Batch Operations: Can touch 10-15 pages in one ingest pass—humans would never do this voluntarily
• Consistent Style: Schema configuration enforces uniform structure and quality across all pages
• Proactive Maintenance: Agents can be scheduled to lint the wiki, finding orphans and contradictions

Key Principles

1. Division of Labor: Human curates sources and asks questions; Large Language Model does everything else (summarizing, filing, cross-referencing, maintenance).

1. Schema-Driven Discipline: The schema document (CLAUDE.md) transforms a generic chatbot into a disciplined wiki maintainer with specific workflows.

1. Autonomous Operation: Agent operates independently on scheduled cycles (e.g., daily ingest at 10 AM, nightly review at 8 PM).

1. Transparency: All operations logged in log.md; humans can review changes via git diff or Obsidian's history.

1. Human Oversight: Humans remain in charge—can reject updates, guide emphasis during ingest, and refine schema over time.

How It Works

Component 1: Schema Configuration

The schema document specifies:

• Directory structure (wiki/raw/, wiki/summaries/, etc.)
• Page templates for each type (summary, entity, concept, synthesis)
• Naming conventions (kebab-case, title casing)
• Quality standards (executive summary length, required sections)
• Workflows (ingest steps, query process, lint checklist)

Component 2: Agent Roles

At retired internal project, different agents have different wiki responsibilities:

Julius (Governance): Primary maintainer. Handles ingest workflow, creates summaries, updates index/log.

Octavius (Knowledge & Effectiveness): Nightly reviewer. Reads all agent files, identifies gaps, forces improvements via team vote.

Cypher (Platform): Infrastructure changes. Updates architecture docs when systems change.

Reacher (Validation): Quality checks. Assesses source credibility and flags low-quality ingests.

Component 3: Scheduled Operations

Daily Ingest Cycle (10 AM):

1. Agent reads new sources from wiki/raw/
2. Creates summary pages following template
3. Updates entity/concept pages mentioned in sources
4. Updates index.md with new/modified pages
5. Appends entries to log.md

Nightly Review Cycle (8 PM):

1. Octavius reviews all agents' work from the day
2. Identifies gaps, inconsistencies, improvement opportunities
3. Files synthesis pages for patterns discovered
4. Reports standup summary to team

Weekly Lint Pass:

1. Health-check for contradictions between pages
2. Find orphan pages (no inbound links)
3. Identify concepts mentioned but lacking pages
4. Flag stale claims superseded by newer sources

Workflow: Ingest Example

Human action: Add karpathy-Large Language Model-wiki-pattern.md to wiki/raw/
     ↓
Julius reads source, identifies key concepts
     ↓
Creates summaries/karpathy-Large Language Model-wiki-pattern.md
     ↓
Updates entities/karpathy-andrej.md (new entity)
     ↓
Updates concepts/persistent-knowledge-base.md (reinforces definition)
     ↓
Updates concepts/agent-maintained-wiki.md (this page—validates pattern)
     ↓
Creates concepts/compounding-artifact.md (new concept)
     ↓
Updates index.md with all new/modified pages
     ↓
Appends to log.md: "## [2026-04-07] ingest | Large Language Model Wiki Pattern"

Implementation at retired internal project

Architecture

Agents interact with wiki through file operations:

┌─────────────┐      ┌──────────────┐      ┌────────────┐
│   Human     │      │  Large Language Model Agent   │      │   Wiki     │
│             │      │              │      │            │
│ Curates     │─────▶│ Reads raw/   │─────▶│ Creates    │
│ sources     │      │ Follows      │      │ summaries/ │
│ Asks        │      │ schema       │      │ Updates    │
│ questions   │◀─────│ Synthesizes  │◀─────│ entities/  │
│             │      │ answers      │      │ concepts/  │
└─────────────┘      └──────────────┘      └────────────┘

Configuration

Agents are configured via:

• CLAUDE.md: Wiki schema and workflows
• Agent-specific instructions: Role-based responsibilities (Julius = governance, Octavius = review)
• Cron jobs: Scheduled operations (daily ingest, nightly standup)

Related Systems

• concepts/retired-internal-project-architecture — Multi-agent system design
• retired-internal-project/julius-governance-agent — Primary wiki maintainer role
• concepts/agent-maintained-wiki — Nightly review and synthesis
• concepts/compounding-artifact — Team coordination mechanism

Governance Decision: Wiki as Infrastructure (2026-05-26)

Decision: The wiki is not a project with an endpoint — it is persistent infrastructure in the agentic stack.

This was formalized in a governance discussion on 2026-05-26 and enacted by updating AGENTS.md and wiki-library-boundary-policy.md with the following rules:

Knowledge retrieval order (fundamental schema):

1. Wiki first (our second brain — interpreted knowledge, decisions, procedures)
2. Library second (canonical source material)
3. Web research only when necessary

Additional rule: Every web research result must immediately produce a new wiki entry so the knowledge base grows and future queries stay internal.

Key framing: The wiki grows forever, in the direction that most benefits our work. It is built for ourselves, as we work — not as a deliverable or milestone.

Benefits

Trade-offs & Limitations

Related Concepts

• concepts/persistent-knowledge-base — The wiki architecture that agents maintain
• concepts/compounding-artifact — Result of sustained agent maintenance
• concepts/ingest-workflow — Primary operation agents perform on sources
• concepts/agent-maintained-wiki — Health-check process for wiki quality
• retired-internal-project/schema-configuration — Document that makes agents disciplined maintainers

Sources & Further Reading

• summaries/karpathy-llm-wiki-pattern — Karpathy's foundational document
• https://gist.GitHub.com/karpathy/442a6bf555914893e9891c11519de94f — Original source

Open Questions

⚠️ Unresolved:

Source excerpt truncated at 220 of 231 lines. Open the canonical wiki path above for the full page.