romfast/claude-plugins
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
cb1e95948e12e489b700f1f0009a1ee41bbf15e1
claude-plugins/CHANGELOG.md
Claude Agent 0207af18f6 feat(learn): Minimalist memory format v3.0 - 60% size reduction 
BREAKING CHANGE: New compact entry format
- Header: `## {P|G}: {Title}` (P=Pattern, G=Gotcha)
- Metadata: `@YYYY-MM-DD #tag1 #tag2 | source:confidence[:reinforced]`
- Max 3 inline tags, no separators, no statistics section

Removed verbose fields: Discovered, Source, Confidence, Trigger
Statistics now calculated dynamically by /learn:status

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2026-01-07 00:44:48 +00:00

4.5 KiB
Raw Blame History

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[3.0.0] - 2026-01-07

Changed (BREAKING)

  • Minimalist memory format: ~60% reduction in file size
    • Old: ### Titlu + 3 metadata lines + separators (~600 char/entry)
    • New: ## P: Titlu + 1 metadata line (~200 char/entry)
  • Header format: ## {P|G}: {Titlu} (P=Pattern, G=Gotcha)
  • Metadata format: @YYYY-MM-DD #tag1 #tag2 | source:confidence[:reinforced]
  • Tags: Inline with #, max 3 per entry
  • Statistics: Removed from files, calculated dynamically by /learn:status

Removed

  • **Discovered**:, **Source**:, **Confidence**:, **Trigger**: fields
  • --- separators between entries
  • ## Domain Statistics section from memory files
  • (feature: ...) context notation

Migration

  • Entries in v2.x format are read normally
  • Auto-converted to v3.0 format on first edit
  • Run /learn:analyze to test new format

[2.1.0] - 2026-01-06

Added

  • Selective saving with signal detection: /learn:analyze now only saves lessons with clear signals
    • Explicit commands: "ține minte", "memorează", "remember", "don't forget"
    • Stated preferences: "întotdeauna", "niciodată", "always", "never", "from now on"
    • Repeated corrections: Same issue corrected 2+ times
    • Historical references: "cum am discutat", "as we discussed"
  • New /learn:cleanup command: Review and remove low-value entries
    • Scoring system based on source, confidence, age, and reinforcement
    • Conservative mode: always confirms before deleting
    • Upgrade option: rewrite entries to make them useful
  • Metadata tracking for entries:
    • source: explicit | inferred | repeated
    • confidence: high | medium | low
    • reinforced: count of times used in subsequent sessions
    • trigger: original phrase that triggered saving
  • Reinforcement tracking: Entries are validated when used in subsequent sessions

Changed

  • /learn:analyze output now shows:
    • Recommended lessons (high signal)
    • Possible lessons (low signal)
    • Excluded one-time fixes
  • Hook changed from Stop to SessionEnd - reminder shows on /exit, /clear, Ctrl+D
  • Templates updated with new metadata fields
  • Statistics now include Total Reinforcements and Last Cleanup

Backward Compatibility

  • Entries without metadata are treated as: source=inferred, confidence=medium, reinforced=0
  • Metadata added automatically on first edit or reinforcement

[2.0.0] - 2026-01-06

Changed (Repository)

  • Reorganized for multi-plugin marketplace: plugin/ → plugins/learn/
  • Repository to be renamed: claude-learn → claude-plugins
  • Updated marketplace.json source path

Added

  • Domain-based memory: Memory is now split into domains (backend, frontend, database, testing, deployment, global)
  • Selective loading via paths frontmatter: Each domain file has a paths: frontmatter that triggers loading only when working with matching files
  • Automatic domain detection: /learn:analyze detects the domain based on file paths in the conversation
  • Automatic pattern suggestion: When creating new domains, glob patterns are auto-generated from conversation files
  • Cross-domain analysis: /learn:reflect now analyzes duplicates and misplaced entries across domains
  • Domain configuration file: .claude/rules/claude-learn-domains.md for user-editable domain definitions
  • New templates: domains-config.md, domain-memory.md

Changed

  • /learn:analyze now prompts for domain confirmation before saving
  • /learn:status aggregates statistics from all domain files with breakdown table
  • /learn:import routes entries to appropriate domains with detection
  • /learn:reflect includes cross-domain duplicate detection and domain migration suggestions
  • Memory files now use format claude-learn-{domain}.md instead of single claude-learn-memory.md

Migration

  • Run /learn:import .claude/rules/claude-learn-memory.md to migrate from v1.0 format
  • Old format is auto-detected and migration is offered

[1.0.0] - 2025-01-06

Added

  • Initial release
  • /learn:analyze - Analyze conversation for patterns and gotchas
  • /learn:import - Import from other memory files
  • /learn:reflect - Consolidate and clean memory
  • /learn:status - Show memory statistics
  • Stop hook for automatic suggestions
  • Support for markdown and JSON import formats