v3.0.0-rc.1 — Full Memory Hierarchy Integration

What's New

Full integration with all 6 Claude Code memory tiers — /reflect now discovers, routes to, and deduplicates across the complete memory hierarchy.

New Memory Targets

Target	Path	Description
Project Rules	.claude/rules/*.md	Modular rules with optional paths: frontmatter for path-scoping
User Rules	~/.claude/rules/*.md	Global modular rules
CLAUDE.local.md	./CLAUDE.local.md	Personal learnings (gitignored)
Auto Memory	~/.claude/projects/<project>/memory/*.md	Low-confidence staging area

New Features

• Hierarchy-aware routing — guardrails to .claude/rules/guardrails.md, model prefs to global, low-confidence to auto memory
• /reflect --organize — analyze memory hierarchy and suggest reorganization (overgrown files, wrong-tier entries, scattered topics, promotion candidates)
• Cross-tier duplicate detection — Step 4 now searches all memory tiers
• Auto memory enrichment — low-confidence items routed to auto memory for later promotion
• Rule file mapping — learning type to suggested rule file with YAML frontmatter guidance

New Utilities (reflect_utils.py)

• _parse_rule_frontmatter() — line-based YAML parser (no PyYAML dependency)
• get_project_folder_name() / get_auto_memory_path() — auto memory path encoding
• read_auto_memory() / suggest_auto_memory_topic() — read and categorize auto memory
• read_all_memory_entries() — cross-tier reader for deduplication
• find_claude_files() — extended with local, rule, user-rule types
• suggest_claude_file(learning_type=) — enhanced routing parameter

Tests

• 197 tests pass (33 new in test_memory_hierarchy.py + 4 backward-compat)
• Tested against 6 real repos with 971 entries across all tiers

Breaking Changes

None — fully backward compatible. find_claude_files() returns new types alongside existing ones. suggest_claude_file() accepts optional learning_type parameter.

---

Full Changelog: v2.6.0...v3.0.0-rc.1

v2.6.0 - Session retention warning

What's New

• Session retention warning — SessionStart hook warns when cleanupPeriodDays is not configured. Claude Code deletes sessions after 30 days by default, which affects /reflect --scan-history and /reflect-skills. Warning disappears once user adds {"cleanupPeriodDays": 99999} to ~/.claude/settings.json.

Full Changelog: v2.5.1...v2.6.0

v2.5.1 - False positive filtering fix

Fixed

• False positive filtering - System content (<task-notification>, <system-reminder>, session continuations) no longer triggers false pattern matches (#15)
  • Added should_include_message() filter before pattern detection
  • Added MAX_CAPTURE_PROMPT_LENGTH (500 chars) guard — real corrections are short, system content is long
  • Explicit remember: markers bypass length filter

Changed

• Made _should_include_message public as should_include_message() (backward-compatible alias preserved)
• Test count increased from 141 to 160

---

Thanks to @DmitryBMsk for the contribution!

v2.5.0 - Session Start Reminder

What's New

Session Start Reminder - When you start a new Claude Code session, if there are pending learnings in the queue, you'll see a helpful reminder.

Features

• Shows up to 5 pending learnings with confidence scores
• Appears at session start — the right time to run /reflect
• Can be disabled via CLAUDE_REFLECT_REMINDER=false environment variable

Contributors

Thanks to @xqliu for this contribution! (#13)

Full Changelog: v2.4.0...v2.5.0

v2.4.0

What's New

• Capture Feedback - Hooks output confirmation when learnings are captured
• Confidence in /view-queue - Shows scores, patterns, relative timestamps
• Guardrail Pattern Detection - Detects "don't do X" constraints with high confidence
• Contradiction Detection - Finds conflicting CLAUDE.md entries in --dedupe

Tests

141 tests passing (15 new tests added)

Full Changelog: v2.3.3...v2.4.0

v2.3.3

Features

• CLAUDE.md size awareness in /reflect --targets:
  • Shows line count for each CLAUDE.md file
  • ✓ indicator for files under 150 lines (healthy)
  • ⚠️ indicator for files over 150 lines (needs cleanup)
  • Auto-recommends --dedupe when threshold exceeded

Fixes

• Cross-platform line counting using Python (Windows support)

Addresses #12

v2.3.2

Fix

• Windows path consistency: Use forward slashes for relative_path on all platforms

Previously, find_claude_files() returned OS-native path separators, causing inconsistent paths on Windows (e.g., ./src\CLAUDE.md instead of ./src/CLAUDE.md).

v2.3.1 - Check existing commands before proposing skills

What's Changed

Bug Fix

• /reflect-skills now checks existing commands before proposing skill candidates
• Shows "Existing skills (patterns match)" section in output
• Prevents proposing duplicates like /campaign-analytics when it already exists

v2.3.0 Changes (included)

• Fixed command directory: ./commands/ → .claude/commands/
• Default to current project (use --all-projects for cross-project scan)
• Added project assignment step - asks where each skill should go
• Added validation step to verify skills work after creation
• Uses existing extraction script for faster processing
• Progress visibility with TodoWrite tracking

Full Changelog: v2.2.1...v2.3.1

v2.2.1 - Bug Fixes

Bug Fixes

Critical: Session Message Extraction

• Fixed extract_user_messages() to handle string content format
• Session files can have {"content": "text"} instead of {"content": [...]}
• This was preventing /reflect-skills from analyzing session history

High: Reduced False Positive Rate

Added structural FALSE_POSITIVE_PATTERNS to reject non-corrections:

• Questions ending with ?
• Task request openers (please help, can you, I need)
• Error descriptions (could not, failed to)
• Bug reports (is not working, is broken)

Improved: Length-Based Confidence

• Short messages (<80 chars) get confidence boost (+0.10)
• Long messages (>150 chars) get reduced confidence
• This helps distinguish direct corrections from context-heavy task requests

Design Philosophy

Pattern detection now uses structural signals for language-agnostic filtering:

• Message length
• Punctuation (question marks)
• Sentence structure patterns

English regex patterns serve as a fast first-pass filter. Non-English corrections are caught by semantic filtering during /reflect.

Tests

• 126 tests passing
• 8 new tests for false positive rejection and string content handling

Full Changelog: v2.2.0...v2.2.1

v2.2.0 - Skill Discovery & Improvement

What's New

Skill Discovery (/reflect-skills)

A new AI-powered command that analyzes your session history to discover repeating patterns that could become reusable skills.

• AI-powered detection — uses reasoning, not regex, to find patterns
• Semantic similarity — detects same intent across different phrasings
• Generates skill files — creates draft skills in ./commands/

Skill Improvement Routing

When you correct Claude while using a skill (e.g., /deploy), the correction can be routed back to the skill file itself:

User: /deploy
Claude: [deploys without running tests]
User: "no, always run tests before deploying"

→ /reflect detects this relates to /deploy
→ Offers to add learning to commands/deploy.md
→ Skill file updated with new step

Subdirectory CLAUDE.md Support (#11)

Now discovers and offers to write to CLAUDE.md files in subdirectories:

• Auto-discovery of ./**/CLAUDE.md files
• Smart target suggestion based on learning content
• Excludes node_modules, .git, venv, etc.

Full Changelog

• feat: Add /reflect-skills command for AI-powered skill discovery
• feat: Add skill improvement routing to /reflect workflow
• feat: Support subdirectory CLAUDE.md discovery (#11)
• test: Add 9 new tests for file discovery functionality
• docs: Update README and SKILL.md for new features

119 tests passing