The Learnings

Passive learning system that trains on your organization habits to improve AI suggestions over time

What is The Learnings?

The Learnings is a passive learning system that observes your organization behavior, corrections, and feedback to continuously improve AI suggestions. It builds a personalized understanding of how you prefer to organize files.

All learning data is encrypted with AES-256, stored locally only, and protected by Touch ID/Face ID.

How It Works

1. Observe

   Sorty passively tracks your organization behavior:

   • Manual corrections (files you move after AI organization)
   • Rejected sessions (organizations you cancel or revert)
   • Custom instructions you provide
   • Steering prompts (post-organization feedback)
2. Analyze

   The system analyzes patterns:

   • Which folder placements you accept vs. reject
   • Temporal weighting (recent behavior > old behavior)
   • Rule induction (deriving explicit rules from patterns)
   • Context-aware preferences (different rules for different file types)
3. Apply

   Learned preferences are injected into future AI prompts:

   • High-priority rules influence organization
   • Rejection patterns prevent repeated mistakes
   • Honing answers define your philosophy
   • Confidence scoring ensures safe application

Learning Sources

The system learns from multiple behavior types:

Behavior	What’s Captured	Priority	Example
Steering Prompts	Post-organization feedback	Highest	”Keep all client files together”
Honing Answers	Explicit preferences from Q&A	High	”Archive files after 6 months”
Guiding Instructions	Pre-organization instructions	High	”Group by client, then by year”
Manual Corrections	Files moved after AI org	Medium	Moving invoice.pdf from Docs/ to Finance/
Reverts	Cancelled organizations	Medium	Cancelling a 20-folder structure
Additional Instructions	Custom org instructions	Medium	”Use snake_case for folders”

Initial Setup

1. Grant Consent

   Navigate to The Learnings (⇧⌘L) and click Grant Consent.

   You’ll be asked to agree to data collection.

2. Set Up Authentication

   Configure Touch ID / Face ID / Passcode protection.

   After initial setup, you’ll need to authenticate each time you access The Learnings dashboard.

3. Start Organizing

   Use Sorty normally. The system will passively observe and learn.

4. Refine with Honing (Optional)

   Periodically run Honing Sessions to explicitly define your preferences.

The Learnings Dashboard

Overview

Quick stats and learning progress:

• Total Corrections: 47
• Honing Answers: 12
• Inferred Rules: 8
• Success Rate: 94%
• Last Updated: 2 hours ago

Action buttons:

• 🧠 Refine Preferences (start honing)
• 📤 Export Profile
• 🚫 Pause Learning
• 🗑️ Delete All Data

Preferences

Grouped view of learned preferences:

Honing Answers:

• When archiving, move to Archives/[Year]/
• Keep projects intact, don’t split by file type
• Use lowercase folder names with hyphens

Inferred Rules:

• .pdf invoices → Finance/Invoices/
• .jpg screenshots → Screenshots/[Year]/
• Client files → Clients/[Name]/

Recent Feedback:

• “Don’t put images in Archives”
• “Keep design files in one folder”

Activity

Timeline of corrections and reverts:

2 hours ago: Moved invoice.pdf
  From: Documents/
  To:   Finance/Invoices/

1 day ago: Cancelled organization
  Reason: Too many folders (15)

3 days ago: Instruction added
  "Group by client name"

Each entry is expandable for details.

Honing Sessions

Honing is an interactive Q&A process to explicitly define your preferences:

1. Trigger Honing

   Click Refine Preferences or use ⌥⌘H.

   Honing is also offered automatically after completing an organization.

2. Answer Questions

   The AI generates 3-5 questions based on your recent activity:

   Example Questions:

   • “When you finish a project, what is your preferred archival strategy?”
   • “How do you prefer to organize documents by date?”
   • “Should design files be grouped by project or by file type?”
3. Provide Answers

   Your answers become high-priority preferences:

   Q: "When archiving old files, where should they go?"
   A: "Move to Archives/[Year]/ and keep folder structure intact."
   
4. Apply to Future Organizations

   These preferences are injected into the AI prompt with high weight.

Prompt Context Generation

The Learnings system generates an XML context block for the AI:

<learnings_context>
  <preamble>
    <instruction>IMPORTANT: The following learnings represent this user's 
    organization preferences, corrections, and patterns. Pay close attention 
    to rejection patterns and corrections - these indicate what NOT to do.</instruction>
    <context folder="Downloads">/Users/me/Downloads</context>
  </preamble>

  <section id="rejections" priority="CRITICAL">
    <title>REJECTION PATTERNS</title>
    <instruction>NEVER use these placements - user explicitly rejected them</instruction>
    <items>
      <item weight="90" occurrences="3" recency="2 days ago">
        DO NOT place .pdf → Archives
      </item>
      <item weight="85" occurrences="2" recency="1 week ago">
        DO NOT place .jpg → Documents
      </item>
    </items>
  </section>

  <section id="preferences" priority="CRITICAL">
    <title>USER PREFERENCES</title>
    <instruction>These are explicit user preferences - always follow them</instruction>
    <items>
      <item weight="98" recency="3 days ago">
        Archive files after 6 months to Archives/[Year]/
      </item>
      <item weight="95" recency="1 week ago">
        Keep all client files in Clients/[Name]/ structure
      </item>
    </items>
  </section>

  <section id="corrections" priority="HIGH">
    <title>PAST CORRECTIONS</title>
    <instruction>User corrected these placements - avoid repeating the same mistakes</instruction>
    <items>
      <item weight="75" occurrences="4" recency="today">
        .pdf files: prefer 'Finance/Invoices/' over 'Documents/'
      </item>
    </items>
  </section>

  <section id="high_confidence_rules" priority="HIGH">
    <title>PROVEN PATTERNS</title>
    <instruction>These patterns have high success rates - follow them unless user instructions conflict. When you apply a pattern, include its rule_id in the folder's JSON.</instruction>
    <items>
      <item weight="85" confidence="92%" recency="2 days ago" rule_id="rule-abc-123">
        .jpg screenshots → Screenshots/[Year]/[Month]/
      </item>
    </items>
  </section>
</learnings_context>

Priority Hierarchy

1. CRITICAL: Rejections, Explicit Preferences (weight 85-100)
2. HIGH: Corrections, High-Confidence Rules (weight 70-84)
3. MEDIUM-HIGH: Recent Feedback (weight 60-69)
4. MEDIUM: Learned Patterns (weight 40-59)
5. LOW: General Tendencies (weight 0-39)

Temporal Weighting

Recent behavior is weighted more heavily than old behavior:

private func recencyWeight(from date: Date, to now: Date) -> Double {
    let daysSince = now.timeIntervalSince(date) / 86400
    if daysSince < 1 { return 1.0 }   // Today: 100%
    if daysSince < 7 { return 0.8 }   // This week: 80%
    if daysSince < 30 { return 0.5 }  // This month: 50%
    if daysSince < 90 { return 0.3 }  // This quarter: 30%
    return 0.1                         // Older: 10%
}

This prevents outdated preferences from affecting current organizations.

Rule Attribution

When the AI applies a learned rule, it includes the rule_id:

{
  "folders": [
    {
      "name": "Finance",
      "rule_id": "rule-invoice-placement-abc123",
      "files": [...]
    }
  ]
}

This enables:

• Feedback loops: “This rule worked well” or “Don’t use this rule again”
• Transparency: See which learned patterns influenced organization
• Debugging: Understand why files were placed where they were

Security & Privacy

Encryption

All learning data is encrypted before storage:

// AES-256 encryption with Keychain-stored keys
try LearningsFileManager.save(profile: currentProfile)

// Keys are never accessible outside the security context
let key = Keychain.load("learnings.encryption.key")

Authentication

Touch ID / Face ID is required after initial setup:

public func unlock() async {
    await securityManager.authenticateForLearningsAccess()
    isLocked = !securityManager.isUnlocked
    if !isLocked {
        await loadProfile()
    }
}

Session Timeout

Automatic lock after 5 minutes of inactivity:

private let sessionTimeout: TimeInterval = 300 // 5 minutes

Secure Deletion

When you delete learning data, it’s securely overwritten:

public func clearAllData() async -> Bool {
    try await consentManager.deleteAllData()
    try LearningsFileManager.secureDelete() // Overwrites before removal
    currentProfile = LearningsProfile() // Reset to empty
    return true
}

Data Management

Pause Learning

Temporarily stop data collection while preserving existing data:

public func withdrawConsent() async {
    consentManager.withdrawConsent()
    profile.consentGranted = false
    await saveProfile()
}

You can resume learning anytime without losing your learned preferences.

Export Profile

Export your learning profile as JSON:

learnings export > my-learnings.json

Example Export:

{
  "createdAt": "2026-01-01T00:00:00Z",
  "honingAnswers": [
    {
      "questionId": "q1",
      "selectedOption": "Archive to Archives/[Year]/"
    }
  ],
  "inferredRules": [
    {
      "id": "rule-123",
      "pattern": ".pdf invoices → Finance/Invoices/",
      "confidenceLevel": "high",
      "successRate": 0.94
    }
  ],
  "corrections": [...],
  "rejections": [...]
}

Import Profile

Restore from a previous export:

1. Settings → Learnings → Import
2. Select your .json file
3. Confirm merge or replace

Importing will overwrite your current learning profile. Export first if needed.

Learning Behavior Types

1. Manual Corrections

Recorded when you move files after AI organization:

public func recordCorrection(originalPath: String, newPath: String) {
    let example = LabeledExample(
        srcPath: originalPath,
        dstPath: newPath,
        action: .edit
    )
    profile.corrections.append(example)
    debouncedSave()
}

2. Steering Prompts

Post-organization feedback:

public func recordSteeringPrompt(
    _ prompt: String,
    folderPath: String?,
    sessionId: String?
) {
    let steeringPrompt = SteeringPrompt(
        prompt: prompt,
        folderPath: folderPath,
        sessionId: sessionId
    )
    profile.steeringPrompts.append(steeringPrompt)
}

3. Cancelled Organizations

Records rich context when you cancel:

public func recordCancelledOrganization(
    folderPath: String,
    fileCount: Int,
    proposedFolderCount: Int,
    instructions: String?,
    stage: String,
    proposedFolderNames: [String]?,
    regenerationCount: Int
) {
    let cancelled = CancelledOrganization(...)
    profile.cancelledOrganizations.append(cancelled)
}

The system learns why you cancelled: too many folders, wrong structure, etc.

4. History Reverts

Records when you undo an organization:

public func recordHistoryRevert(
    entryId: String,
    operationCount: Int,
    folderPath: String?,
    revertReason: String?
) {
    let event = RevertEvent(...)
    profile.historyReverts.append(event)
}

Auto-Inference

The system can automatically infer rules from patterns:

private func checkAndTriggerAutoInference() async {
    // After 5+ corrections, try to infer a rule
    if profile.corrections.count >= 5 {
        await analyzer.inferRules(from: profile)
    }
}

Inference Example

Observations:

• User moved invoice_jan.pdf from Documents/ to Finance/Invoices/
• User moved invoice_feb.pdf from Documents/ to Finance/Invoices/
• User moved receipt_jan.pdf from Documents/ to Finance/Receipts/

Inferred Rule:

Pattern: Files matching "invoice_*.pdf"
Destination: Finance/Invoices/
Confidence: High (100% success rate)

CLI Commands

Learnings CLI

Manage your learning profile from the terminal:

# Show learning status
learnings status

# Show detailed statistics
learnings stats

# Export profile data
learnings export > my-profile.json

# Delete all learning data
learnings clear

# Pause learning
learnings withdraw

# Show system information
learnings info

Sorty CLI Integration

# Open Learnings dashboard
sorty learnings

# Start honing session
sorty learnings --honing

# View stats
sorty learnings --stats

Deeplinks

Deeplink	Description
sorty://learnings	Open Learnings dashboard
sorty://learnings?action=honing	Start a honing session
sorty://learnings?action=stats	View learning statistics

Related Features

File Organization

Learn how AI organization works

Personas

Customize organization behavior

Workspace Health

Monitor directory health and clutter