user-memory - SKILL.md Agent Skill

name: user-memory description: Detects and persists user-provided personal information across Cowork sessions. Auto-activated when user states preferences, financial situation, or corrections.

User Memory

Purpose

Automatically detect when users provide information about their financial situation, preferences, or corrections, and persist this to the appropriate learnings file so future sessions can use it.

When to Activate

This skill activates automatically when:

User answers a question about their financial situation
User volunteers personal context information
User corrects a categorization, subscription status, or merchant name
User states an output preference

Trigger Patterns

Financial Situation (Danish)

Pattern	Category	Extraction
`Jeg har [X] kr i nødopsparing/opsparing`	emergency_fund	amount
`Min nødopsparing er [X] kr`	emergency_fund	amount
`Jeg har [X] måneder i opsparing`	emergency_fund	months_coverage
`Jeg har gæld i/til [X]`	debt	creditor
`Min gæld er [X] kr`	debt	amount
`Jeg skylder [X] kr til [Y]`	debt	amount, creditor
`Min pension er [X]%`	pension	percentage
`Jeg indbetaler [X]% til pension`	pension	percentage
`Jeg har pension hos [X]`	pension	provider
`Min løn er [X] kr`	income	amount
`Jeg tjener [X] kr om måneden`	income	amount
`Min bruttoløn er [X] kr`	income	gross_amount

Financial Situation (English)

Pattern	Category	Extraction
`I have [X] in emergency fund/savings`	emergency_fund	amount
`My emergency fund is [X]`	emergency_fund	amount
`I have [X] months of expenses saved`	emergency_fund	months_coverage
`I have debt with/to [X]`	debt	creditor
`I owe [X] to [Y]`	debt	amount, creditor
`My pension is [X]%`	pension	percentage
`I contribute [X]% to pension`	pension	percentage
`My salary is [X]`	income	amount
`I earn [X] per month`	income	amount

Personal Context (Danish)

Pattern	Category	Extraction
`Jeg arbejder hos [X]`	employer	company_name
`Min arbejdsgiver er [X]`	employer	company_name
`Jeg er ansat hos [X]`	employer	company_name
`Jeg vil gerne [X]`	goal	goal_description
`Mit mål er at [X]`	goal	goal_description
`Jeg sparer op til [X]`	goal	savings_goal
`Jeg planlægger at [X]`	goal	plan_description
`Jeg er [X]` (single/married/etc.)	life_situation	status
`Jeg har [X] børn`	life_situation	children_count
`Jeg bor i [X]`	life_situation	location
`Jeg er [X] år`	life_situation	age

Personal Context (English)

Pattern	Category	Extraction
`I work at [X]`	employer	company_name
`My employer is [X]`	employer	company_name
`I want to [X]`	goal	goal_description
`My goal is to [X]`	goal	goal_description
`I'm saving for [X]`	goal	savings_goal
`I plan to [X]`	goal	plan_description
`I am [X]` (single/married/etc.)	life_situation	status
`I have [X] children`	life_situation	children_count

Corrections (Danish)

Pattern	Category	Target File
`[X] er ikke [Y], det er [Z]`	category_correction	learnings/categorization.md
`Nej, det er [X], ikke [Y]`	category_correction	learnings/categorization.md
`Det er (ikke) et abonnement`	subscription_correction	learnings/subscriptions.md
`[X] er et abonnement`	subscription_confirmation	learnings/subscriptions.md
`[X] er det samme som [Y]`	merchant_alias	learnings/merchants.md
`[X] hedder egentlig [Y]`	merchant_alias	learnings/merchants.md

Output Preferences (Danish)

Pattern	Category	Extraction
`Vis altid [X]`	display_preference	format
`Jeg foretrækker [X]`	display_preference	preference
`Brug [X] format`	display_preference	format
`Tal dansk/engelsk`	language_preference	language
`Mere/mindre detaljer`	detail_preference	level

Output Preferences (English)

Pattern	Category	Extraction
`Always show [X]`	display_preference	format
`I prefer [X]`	display_preference	preference
`Use [X] format`	display_preference	format
`More/less detail`	detail_preference	level

Target Files

Learning Type	Target File	Section
Emergency fund	learnings/preferences.md	Financial Situation
Debt status	learnings/preferences.md	Financial Situation
Pension status	learnings/preferences.md	Financial Situation
Income	learnings/preferences.md	Financial Situation
Employer	learnings/preferences.md	Personal Context
Goals	learnings/preferences.md	Personal Context
Life situation	learnings/preferences.md	Personal Context
Category corrections	learnings/categorization.md	Learnings table
Subscription status	learnings/subscriptions.md	Confirmed/Not sections
Merchant aliases	learnings/merchants.md	Aliases table
Output preferences	learnings/preferences.md	Output Preferences

Learning Entry Format

For preferences.md

### {YYYY-MM-DD}
**Category**: {financial_situation|personal_context|output_preference}
**Trigger**: "{exact user statement}"
**Rule**: {extracted structured information}
**Source**: {command that captured this}

For categorization.md

Append to the Learnings table:

| {YYYY-MM-DD} | {pattern} | {merchant} | {category} | {subcategory} | {user statement} |

For subscriptions.md

Append to Confirmed Subscriptions:

| {YYYY-MM-DD} | {pattern} | {merchant} | {frequency} | User confirmed |

Or append to Not Subscriptions:

| {YYYY-MM-DD} | {pattern} | {merchant} | User: "{reason}" |

For merchants.md

Append to Aliases table:

| {YYYY-MM-DD} | {raw_pattern} | {normalized_name} | User alias |

Workflow

When user provides information that matches a trigger pattern:

Detect: Identify which pattern matched
Extract: Pull out the relevant values (amounts, names, etc.)
Validate: Ensure the extracted information makes sense
Target: Determine which file and section to write to
Check duplicates: Read target file, check if similar rule exists
Write: Append the learning entry in the correct format
Confirm: Acknowledge briefly without interrupting flow

Confirmation Behavior

After saving a learning, acknowledge briefly in Danish:

Learning Type	Confirmation
Financial status	"Noteret: {summary}. Jeg husker det."
Personal context	"Noteret: {summary}."
Correction	"Forstået. Jeg retter det fremover."
Preference	"Noteret. Jeg bruger det fremover."

Rules:

Keep confirmations short (one sentence)
Don't interrupt the conversation flow
Integrate naturally into the response
Don't ask for confirmation before saving

Reading Learnings

At the start of relevant workflows, read learnings to apply prior knowledge:

Before asking about emergency fund: Check preferences.md for existing emergency_fund entry
Before asking about pension: Check preferences.md for existing pension entry
Before asking about employer: Check preferences.md for existing employer entry
Before categorizing: Check categorization.md for learned rules
Before detecting subscriptions: Check subscriptions.md for confirmed/denied

If a learning exists, skip the question and use the stored value.

Update Behavior

When user provides updated information:

Detect update: User provides new value for existing learning
Mark old entry: Add [Superseded by {date}] to old entry
Write new entry: Append new learning with current date
Confirm: "Opdateret: {old_value} → {new_value}"

Example: Advice Session

User: /smartspender:advice

Claude: [Reads learnings/preferences.md - no emergency_fund entry]
Claude: "Har du en nødopsparing?"

User: "Ja, jeg har 20.000 kr i opsparing"

Claude: [Detects pattern: "har X kr i opsparing"]
Claude: [Extracts: amount=20000, category=emergency_fund]
Claude: [Writes to learnings/preferences.md]
Claude: "Noteret: 20.000 kr i nødopsparing. [continues with advice]"

---
Next session:

User: /smartspender:advice

Claude: [Reads learnings/preferences.md - finds emergency_fund=20000]
Claude: [Skips emergency fund question, uses stored value]
Claude: "Baseret på din nødopsparing på 20.000 kr..."

Example: Category Correction

User: "Netflix er underholdning, ikke abonnement"

Claude: [Detects pattern: "X er Y, ikke Z"]
Claude: [Extracts: merchant=Netflix, correct=Underholdning, incorrect=Abonnementer]
Claude: [Writes to learnings/categorization.md]
Claude: "Forstået. Netflix kategoriseres som Underholdning fremover."

Integration Points

Commands that should trigger user-memory:

/smartspender:advice - Financial questions
/smartspender:analyze - Category corrections
/smartspender:subscriptions - Subscription confirmations
/smartspender:overview - Preference statements
Any conversation where user volunteers information

Conflict Resolution

If stored learning conflicts with new information:

Prefer recent: Newer information supersedes older
Ask if ambiguous: "Tidligere sagde du X, nu siger du Y. Hvad er korrekt?"
Update on confirmation: Write new entry, mark old as superseded