memgrafter/knowledge-registry

Register specific files or line ranges as curated knowledge — not the whole codebase, that's what git is for. Then retrieve from them — search with rg or dump content — to build focused agent context.

Why use it

• Retrieve from multiple files/ranges in one command instead of repeating paths
• Search across registered sources with kr search <reg> "query"
• Dump all content to stdout for piping into agent prompts
• Line-range support — register only the relevant sections of large files

When to use

• You need to retrieve information from a curated set of files — search or dump as a unit
• Building focused context for a coding task by retrieving across multiple registered sources
• Searching within a subset of a codebase instead of the whole repo

When NOT to use

• You need the whole codebase — that's what git is for; register only curated knowledge, not every file
• One-off single-file searches — just use rg directly
• You need remote/URL sources — only local filesystem supported now

Usage

bash
kr registry create <name> -d "Purpose of this KR"   # always add a description
kr source add <reg> file:///path/to/file.rs#L1-L50 --label "main" --tags "core"
kr source list <reg>
kr search <reg> "query" -c 2          # rg with context lines
kr dump <reg>                         # all content to stdout
kr registry update <name> -d "Updated purpose"  # update description
kr source update <reg> 0 --label "new" --tags "updated"
kr source remove <reg> 0
kr registry delete <name>

URI formats: file:///path.rs (full file), file:///path.rs#L10-L42 (range), file:///path.rs#L10 (from line 10 to end).

Path resolution: URIs are stored relative to the .kr folder parent. At display time, paths resolve to ~/... if under home, /... otherwise — no brittle absolute paths in the registry file.

Cross-registration: A source can belong to multiple KRs if it's relevant to both domains. Overlap is intentional cross-referencing, not duplication.

Examples

bash
# Register curated knowledge for an auth task — specific sections, not the whole repo
kr registry create auth-knowledge -d "Auth module: handler code + design notes"
kr source add auth-knowledge file:///path/to/app/src/auth.rs#L1-L80 --label "auth module"
kr source add auth-knowledge file:///path/to/app/DESIGN.md#L15-L30 --label "auth design notes"
kr search auth-knowledge "fn authenticate"
kr dump auth-knowledge > context.md

Output

• search: rg results grouped by source with headers
• dump: line-numbered content per source with URI/label headers

Description: Every KR should have a description (set with -d on create or update). This helps the model recall the purpose of a KR when it returns to it in a later session.

Cost: local only, zero network. Benefit: clean context management across multi-file tasks.