lore

Lore — Domain Knowledge Plugin for Claude Code

A Claude Code plugin that creates and maintains a self-correcting domain knowledge base per project. Business rules, entity definitions, architectural decisions, and domain context persist across sessions — so agents stop re-researching the same rules and stop making the same wrong assumptions.

How it works

Knowledge lives in a lore/ directory in your project repo — version-controlled, PR-reviewable, team-shared. Each entry is an atomic markdown file with YAML frontmatter tracking confidence level, sources, and domain classification. Entries connect to each other through [[wikilinks]], forming a knowledge graph that agents can traverse.

Agents check the lore before asking you domain questions. When they learn something new, they write it back. When they get something wrong, a retrospective hook catches it and the entry gets corrected.

Confidence lifecycle

Every entry carries a confidence level that reflects how it was validated:

assumed  →  inferred  →  verified

• assumed — Unverified agent inference. Treat with caution.
• inferred — Derived from code analysis or research. Likely correct but not human-confirmed.
• verified — Human-confirmed. The ground truth.

Only human confirmation promotes confidence. Agents always present lore-sourced context as explicit assumptions you can correct.

Knowledge graph

Entries link to each other using [[wikilinks]]. When you query the lore, the search engine uses these links to surface related context:

• BM25 ranking scores results by keyword relevance, not naive string matching.
• Inbound link boost promotes well-connected hub pages — central concepts surface first.
• 1-hop expansion automatically includes linked neighbor pages from top results, interleaved by score.
• Confidence bonus ranks verified entries above inferred and assumed ones.
• Recency bonus gives a small boost to recently updated entries.

Source provenance

Original documents (CLAUDE.md, internal docs, READMEs) are preserved as source-type pages during bootstrap. Every derived entry traces back to its source page via the sources field, and source pages track their derivatives via derived_entries. This creates a full provenance chain — you can always trace any claim back to where it came from.

Quick start

1. Install the plugin

/plugin marketplace add leviosa-medical/lore
/plugin install lore@lore

2. Set up a project

From your target project's root:

/lore:init

This creates:

• lore/ directory with subdirectories for each entry type
• lore/index.md — entry catalog with wikilinks
• lore/config.yaml — project-specific domains and routing guidance
• lore/operations.jsonl — append-only operation log
• .claude/rules/lore.md — agent integration rule (makes agents query lore automatically)
• .claude/settings.json — marketplace and plugin config (so teammates get auto-prompted to install)

Important: Restart Claude Code after running /lore:init. The MCP server and agent rule aren't available until after a restart.

3. Seed initial knowledge

/lore:bootstrap

This reads your CLAUDE.md, .claude/rules/, skill definitions, and docs to extract declarative domain knowledge into atomic, wikilinked lore entries. Imperative instructions ("always do X", "never do Y") stay in CLAUDE.md. After seeding, the skill suggests which declarative lines can be removed from CLAUDE.md to save context tokens.

Usage

Automatic (during agent work)

Once installed, the agent rule in .claude/rules/lore.md instructs agents to:

1. Query first — call lore_query before asking you domain questions, even if the answer seems obvious from files already in context
2. State assumptions — present lore-sourced context as explicit assumptions: "Based on the lore, I'm assuming X. Flag if wrong."
3. Write back — when the lore has no answer, research it and write it back with appropriate confidence

You don't need to invoke any tools manually — the agent does this on its own.

Skills

Skill	Purpose
/lore:init	Set up lore in the current project
/lore:bootstrap	One-time seeding from CLAUDE.md, rules, skill definitions, and docs into atomic wikilinked entries
/lore:capture	Capture domain knowledge that surfaces during conversation
/lore:resolve	Resolve git merge conflicts and semantic contradictions in lore entries

/lore:init

Creates the lore/ directory structure, the agent integration rule, and configures project settings so teammates are auto-prompted to install the plugin. Run once per project.

/lore:bootstrap

Reads existing project knowledge sources and decomposes them into atomic entries: