---

name: reflect description: Intelligent learning system that captures corrections, preferences, and patterns from sessions, saves them to Forgetful Memory, and synthesizes learnings into skill improvements. Use when user says "reflect", "remember this", "learn from this", after corrections, or at session end.

Reflect - Intelligent Learning System

Overview

Reflect captures learnings from conversations and saves them to Forgetful Memory, creating a feedback loop that improves future interactions. Learnings are categorized, linked to skills/projects/entities, and periodically synthesized into skill improvements.

Quick Reference

Command	Purpose
/reflect	Analyze current session, capture learnings
/reflect [skill]	Focus analysis on specific skill
/reflect status	Show learning statistics
/reflect synthesize [skill]	Propose skill improvements from learnings
/reflect graph [skill]	Visualize learning connections
/reflect stale	Find outdated learnings
/reflect conflicts	Find contradictory learnings

Learning Types

See LEARNING_TYPES.md for detailed reference.

Type	Tag	Importance	Signal
Correction	correction	9	User said "no", "not like that", explicit fix
Preference	preference	8	Repeated choices, "I prefer", "always use"
Pattern	pattern	7	Successful approach, "this worked well"
Edge Case	edge-case	7	Unexpected scenario, workaround needed
Anti-pattern	anti-pattern	9	"Never do this", caused problems

Core Workflow

Step 1: Analyze Session

Scan the conversation for learning signals:

HIGH confidence signals (corrections):

• User said "no", "wrong", "not like that", "I meant..."
• User explicitly corrected output
• User requested changes immediately after generation
• User expressed frustration

MEDIUM confidence signals (preferences/patterns):

• User said "perfect", "great", "yes", "exactly"
• User accepted output without modification
• User built upon the output
• Repeated choices across the session

MEDIUM confidence signals (edge cases):

• Questions the approach didn't anticipate
• Scenarios requiring workarounds
• Features user asked for that weren't covered

Step 2: Check for Existing Similar Learnings

Before creating new memories, ALWAYS query for similar existing ones:

# MCP tool invocation pseudocode (not executable Python)
mcp__forgetful__execute_forgetful_tool("query_memory", {
  "query": "<learning topic keywords>",
  "query_context": "Checking for existing similar learnings before creating new",
  "tags": ["skill-learning"],
  "k": 5,
  "include_links": True
})

If similar learning exists:

• Same topic, same conclusion → Update existing (bump importance if repeated)
• Same topic, different conclusion → Flag as conflict, link both
• Related but distinct → Create new, link to existing

Step 3: Determine Scope

Identify what scope this learning applies to:

Scope	When	How to Tag
Personal	User-specific preference	Link to user entity
Project	Project-specific pattern	Use project_ids
Skill	Skill-specific learning	Tag with skill name
Organization	Applies to all Compass Brand	Tag with org-wide

Step 4: Create Structured Memory

mcp__forgetful__execute_forgetful_tool("create_memory", {
  "title": "[TYPE] Brief description",
  "content": "Detailed learning with context. What happened, why it matters, how to apply.",
  "context": "Captured during [skill/task] - [why this is important]",
  "keywords": ["topic1", "topic2", "skill-name"],
  "tags": ["skill-learning", "<type>", "<skill-name>"],
  "importance": <6-10 based on type and confidence>,
  "project_ids": [<if project-specific>]
})

Step 5: Link Related Entities

If the learning involves specific people, technologies, or concepts:

# Link to entity (person, technology, etc.)
mcp__forgetful__execute_forgetful_tool("link_entity_to_memory", {
  "entity_id": <entity_id>,
  "memory_id": <new_memory_id>
})

# Link to related memories
mcp__forgetful__execute_forgetful_tool("link_memories", {
  "memory_id": <new_memory_id>,
  "related_ids": [<related_memory_ids>]
})

Step 6: Present Summary

┌─ Learning Captured ───────────────────────────────────────┐
│                                                           │
│  Type: [CORRECTION/PREFERENCE/PATTERN/etc.]               │
│  Skill: [skill-name or "general"]                         │
│  Scope: [personal/project/skill/org-wide]                 │
│                                                           │
│  Title: [learning title]                                  │
│  Confidence: [HIGH/MEDIUM/LOW] (occurrence #X)            │
│                                                           │
│  Linked to:                                               │
│  • Memory #42: [related learning]                         │
│  • Entity: [person/technology]                            │
│                                                           │
│  Memory ID: #[new_id]                                     │
└───────────────────────────────────────────────────────────┘

Confidence Tracking

Track learning confidence through repetition:

Occurrences	Confidence	Importance	Action
1	Low	6-7	Create memory
2	Medium	7-8	Link to first, note repetition
3+	High	8-9	Consider graduating to skill

When a pattern repeats:

# Update existing memory
mcp__forgetful__execute_forgetful_tool("update_memory", {
  "memory_id": <existing_id>,
  "importance": min(10, current_importance + 1),
  "content": "<updated content noting repetition count>"
})

Integration Points

Proactive Loading

When a skill starts, query relevant learnings:

mcp__forgetful__execute_forgetful_tool("query_memory", {
  "query": "<skill-name> preferences corrections patterns",
  "query_context": "Loading learnings before skill execution",
  "tags": ["skill-learning"],
  "k": 5
})

Correction Detection

Note: Stop hooks were removed from the reflect system (see .claude/settings.json). The system no longer automatically monitors for corrections. Use /reflect manually after sessions or when you want to capture learnings.

Additional Documentation

Document	Purpose
LEARNING_TYPES.md	Detailed reference for all learning types
SYNTHESIS.md	How to graduate learnings to skills
LIFECYCLE.md	Status, stale, and conflict management
VISUALIZATION.md	Graph visualization features
PROACTIVE_LOADING.md	Pre-skill context injection
HOOKS.md	Hook configuration for automation

Commands Reference

Command	Description	Details
/reflect	Analyze session, capture learnings	reflect.md
/reflect status	Learning system health	reflect-status.md
/reflect synthesize	Graduate learnings to skills	reflect-synthesize.md
/reflect graph	Visualize connections	reflect-graph.md
/reflect stale	Find outdated learnings	reflect-stale.md
/reflect conflicts	Find contradictions	reflect-conflicts.md