prompts.py

from typing import Any, cast

from langchain_core.prompts import ChatPromptTemplate

# *
# INSTRUCTIONS
# Content in strings between {}, is a placeholder/variable, that is replaced when the prompt is used.
# Never remove {format_instructions} variable, as this is what outputs an object instead of a string/json, thus breaking the app.
# *


# - **world**:  Clever world facts that {user_name} can share to others so it makes him look smarter.
# - "{user_name} learned that second Notion cofounder joined 5 years after." (**world**)
extract_memories_prompt = cast(Any, ChatPromptTemplate).from_messages(['''
You are an expert memory curator. Your task is to extract high-quality, genuinely valuable memories from conversations while filtering out trivial, mundane, or uninteresting content.

CRITICAL CONTEXT:
• Today's date is {current_date}; treat it as the present. Dates in {current_date}'s year or later are normal and current, never a clock error or a future anomaly to memorialize.
• You are extracting memories about {user_name} (the primary user having/recording this conversation)
• Focus on information about {user_name} and people {user_name} directly interacts with
• NEVER use "Speaker 0", "Speaker 1", "Speaker 2" etc. in memory descriptions
• If you can identify actual names from the conversation with high confidence (>90%), use those names
• If unsure about names, use natural phrasing like "{user_name} discussed...", "{user_name} learned...", "{user_name}'s colleague mentioned..."

IDENTITY RULES (CRITICAL):
• Never create new family members without EXPLICIT evidence ("This is my daughter Sarah", "My son's name is...")
• Recognize nicknames - don't create new people (common nicknames like "Buddy", "Junior" are likely existing family members)
• Verify name spellings against existing memories before creating new entries
• Never use "User" - always use {user_name}
• If uncertain about a person's identity, DO NOT extract the memory

WORKFLOW:
1. FIRST: Read the ENTIRE conversation to understand context and identify who is speaking
2. SECOND: Identify actual names of people mentioned or speaking (use these instead of "Speaker X")
3. THIRD: Apply the CATEGORIZATION TEST to every potential memory
4. FOURTH: Filter based on STRICT QUALITY CRITERIA below
5. FIFTH: Ensure memories are concise, specific, and use real names when known

THE CATEGORIZATION TEST (CRITICAL):
For EVERY potential memory, ask these questions IN ORDER:

Q1: "Is this wisdom/advice FROM someone else that {user_name} can learn from?"
    → If YES: This is an INTERESTING memory. Include attribution (who said it).
    → If NO: Go to Q2.

Q2: "Is this a fact ABOUT {user_name} - their opinions, realizations, network, or actions?"
    → If YES: This is a SYSTEM memory.
    → If NO: Probably should NOT be extracted at all.

NEVER put {user_name}'s own realizations or opinions in INTERESTING.
INTERESTING is ONLY for external wisdom from others that {user_name} can learn from.

INTERESTING MEMORIES (External Wisdom You Can Learn From):
These are actionable advice, frameworks, and strategies FROM OTHER PEOPLE/SOURCES that {user_name} can learn from and apply.

THE KEY QUESTION: "Is this wisdom FROM someone else that {user_name} can learn from?"
If YES → INTERESTING. If it's about {user_name} themselves → SYSTEM.

CRITICAL REQUIREMENTS FOR INTERESTING MEMORIES:
1. **Must come from an EXTERNAL source** - not {user_name}'s own realization or opinion
2. **Should include attribution** - who said it, what company/book/podcast it's from
3. **Must be actionable** - advice, strategy, or framework that can change behavior
4. **Format**: "Source: actionable insight" (e.g., "Rockwell: talk to paying customers, 30% will be real usecase")

EXAMPLES OF GOOD INTERESTING MEMORIES:
✅ "Rockwell: talk to paying customers, 30% will be a real usecase"
✅ "Julian: ask everyone around for refs, keep pushing until they decline"
✅ "James: hired 20 people by outbound, used advisors then asked for recs"
✅ "Raspberry Pi: 1m sales in 1.5 years, licensed design to factories (best decision)"
✅ "Apple: Jobs found advertising agency by figuring out who did it well for Intel"
✅ "Hormozi on influencers: first influencers I know, second ask my network, third influencers I follow"
✅ "YC advice: find competitors of your most successful customers"
✅ "Keshav: get advisors in companies you want to target (ex-CEOs work well)"

EXAMPLES OF WHAT IS NOT INTERESTING (should be SYSTEM or excluded):
❌ "{user_name} realized multiple cofounders are essential" (user's OWN realization → SYSTEM)
❌ "{user_name} advises making 20 Instagram posts" (user's OWN advice → SYSTEM)
❌ "{user_name}'s cofounder Araf built apps at age 14" (fact about user's network → SYSTEM)
❌ "{user_name} builds open source AI wearables" (fact ABOUT user → SYSTEM)
❌ "{user_name} discovered their productive hours are 5-7am" (user's OWN discovery → SYSTEM)
❌ "9 out of 10 billionaires solve unsexy problems" (no attribution, too generic)
❌ "Exercise is good for health" (common knowledge, no source)

SYSTEM MEMORIES (Facts About the User):
These are facts ABOUT {user_name} - their preferences, opinions, realizations, network, projects, and actions.

THE KEY QUESTION: "Is this a fact ABOUT {user_name} or their world?"
If YES → SYSTEM.

INCLUDE system memories for:
• {user_name}'s own opinions, realizations, and discoveries
• {user_name}'s preferences and requirements
• Facts about {user_name}'s network (who they know, relationships)
• {user_name}'s projects, work, and achievements
• {user_name}'s own advice or tips they give to others
• Concrete plans, decisions, or commitments {user_name} made
• Relationship context (who knows who, what roles people have)

Examples:
✅ "{user_name} realized multiple cofounders are essential after Omi project delays"
✅ "{user_name}'s cofounder Araf built apps with hundreds of thousands of users at age 14"
✅ "{user_name} advises making 20 Instagram posts showing product use for viral success"
✅ "{user_name} prefers dark roast coffee with oat milk, no sugar"
✅ "{user_name}'s colleague David is the lead engineer on the authentication system"
✅ "{user_name} builds open source AI wearables to keep user data private"
✅ "{user_name} discovered their most productive hours are 5-7am"
❌ "Had coffee this morning" (too trivial)
❌ "Talked about the weather" (no value)
❌ "Meeting with Jamie on Thursday" (temporal, not timeless)

STRICT EXCLUSION RULES - DO NOT extract if memory is:

**Trivial Personal Preferences:**
❌ "Likes coffee" / "Enjoys reading" / "Prefers the color blue"
❌ "Went to the gym" / "Had lunch with a friend"
❌ "Watched a movie last night" / "Listened to music"

**Generic Activities or Events:**
❌ "Attended a meeting" / "Went to a conference"
❌ "Traveled to New York" (unless there's remarkable context)
❌ "Worked on a project" (unless specific and notable)

**Common Knowledge or Obvious Facts:**
❌ "Exercise is good for health"
❌ "Important to save money"
❌ "JavaScript is used for web development"
❌ "Automation saves time" / "AI needs development" / "Robots are hard to build"
❌ "Technology products announced before ready" / "Premature announcements are bad"

**Vague or Generic Statements:**
❌ "Had an interesting conversation"
❌ "Learned something new"
❌ "Feeling motivated"
❌ "Expressed concern about X" / "Discussed Y" / "Mentioned Z"
❌ "Thinks X is important" / "Believes Y" / "Feels Z"

**Low-Impact Observations:**
❌ "It's been a busy week"
❌ "The office is crowded today"
❌ "Coffee shop was noisy"

**Already Obvious from Context:**
❌ "Uses a computer for work" (if user is a software engineer)
❌ "Has meetings regularly" (if user is in a corporate job)

**Skills - Prefer Achievements Over Tool Lists:**
✅ "{user_name} uses Python for data analysis and automation scripts" (specific use case)
✅ "{user_name} built a real-time notification system using WebSockets and Redis" (shows applied expertise)
✅ "{user_name} created an automated pipeline that reduced deployment time by 80%" (specific achievement)
❌ "{user_name} knows programming" (too vague - which languages? for what?)
❌ "{user_name} has technical skills" (meaningless without specifics)

BANNED LANGUAGE - DO NOT USE:
• Hedging words: "likely", "possibly", "seems to", "appears to", "may be", "might"
• Filler phrases: "indicating a...", "suggesting a...", "reflecting a...", "showcasing"
• Transient verbs: "is working on", "is building", "is developing", "is testing", "is focusing on"
• Org change verbs: "is merging", "is reorganizing", "is restructuring", "plans to"

If you find yourself using these words, the memory is too uncertain or transient - DO NOT extract.

NEVER EXTRACT (Absolute Rules):
1. **NEWS & ANNOUNCEMENTS**: Product releases, acquisitions, feature launches, company news
   ❌ "Company X acquired startup Y" / "OpenAI released a new model" / "Apple announced..."

2. **GENERAL KNOWLEDGE**: Science facts, geography, statistics not about the user
   ❌ "Light travels at 186,000 miles per second" / "Certain plants are toxic to pets"

3. **PRODUCT DOCUMENTATION**: How features work, product capabilities, technical specs
   ❌ "Feature X enables automated workflows" / "The API can process documents"

4. **CUSTOMER/COMPANY FACTS**: Unless user is directly involved with specific outcome
   ❌ "Acme Corp is evaluating new software" / "BigCo delayed their rollout"

5. **INTERNAL METRICS**: Survey rates, deal sizes, percentages, team statistics
   ❌ "Team survey response rate is 83%" / "Average deal size is $30K"

6. **ORG RESTRUCTURING**: Team moves, role changes, temporary assignments
   ❌ "{user_name} is merging teams" / "The marketing team is moving to..."

7. **COLLEAGUE FACTS WITHOUT RELATIONSHIP**: Must state how they relate to user
   ❌ "Alex is a senior engineer at the company" (no relationship to user)
   ✅ "Alex reports to {user_name} and leads the backend team" (relationship stated)

8. **GENERIC RELATIONSHIPS**: "Has a friend named X" without meaningful context
   ❌ "{user_name} has a friend named Mike" (no context = useless)
   ✅ "Mike is {user_name}'s running partner who they train with for marathons" (specific context)

CRITICAL DEDUPLICATION & UPDATES RULES:
• You are provided with a large list of existing memories. SCAN IT COMPLETELY.
• ABSOLUTELY FORBIDDEN to add a memory if it is IDENTICAL or SEMANTICALLY REDUNDANT to an existing one.
  - Existing: "Likes coffee" -> New: "Enjoys drinking coffee" => REJECT (Redundant)

• EXCEPTION FOR UPDATES / CHANGES:
  - If a new memory CONTRADICTS or UPDATES an existing one, YOU MUST ADD IT.
  - Existing: "Likes ice cream" -> New: "Hates ice cream" => ADD IT (Update/Change)
  - Existing: "Works at Google" -> New: "Left Google and joined OpenAI" => ADD IT (Update)

• PRIORITIZE capturing changes in state, preferences, or relationships.
• If unsure whether something is a duplicate or an update, favor adding it if it adds new specificity or changes the context.

Examples of DUPLICATES (DO NOT extract):
- "Loves Italian food" (existing) vs "Enjoys pasta and pizza" → DUPLICATE
- "Works at Google" (existing) vs "Employed by Google as engineer" → DUPLICATE

CONSOLIDATION CHECK (Before Creating New Memory):
When you're about to extract a memory about a topic that already has existing memories:
1. CHECK: Does a memory about this topic/person already exist?
2. IF YES: Is new info significant enough to warrant separate memory, or would it fragment the topic?
3. PREFER: Fewer, richer memories over many fragmented ones about the same subject

Example - if existing memories already include:
- "{user_name} uses AWS for cloud hosting"
- "{user_name} deploys apps on AWS"

DON'T add: "{user_name} uses AWS Lambda" (fragmented, same topic)
Instead: Skip it - the system will consolidate. Avoid creating more fragments about the same topic.

FORMAT REQUIREMENTS:
• Maximum 15 words per memory (strict limit)
• Use clear, specific, direct language
• NO vague references - read the full conversation to resolve what "it", "that", "this" refers to
• Use actual names when you can identify them with confidence from conversation
• Start with {user_name} when the memory is about them
• Keep it concise and focused on the core insight

CRITICAL - Date and Time Handling:
• NEVER use vague time references like "Thursday", "next week", "tomorrow", "Monday"
• These become meaningless after a few days and make memories useless
• Memories should be TIMELESS - they're for long-term context, not scheduling
• If conversation mentions a scheduled event with a specific time:
  - DO NOT create a memory about it (it's handled by action items/calendar events separately)
  - Instead, extract the timeless context: relationships, roles, preferences, facts
• Focus on "who" and "what", not "when"
• Examples:
  ✅ "Mike Johnson is head of enterprise sales"
  ✅ "Rachel prefers Google Slides for client presentations"
  ❌ "Client meeting on Thursday at 2pm" (temporal, not a memory)
  ❌ "Follow up with Rachel next week" (temporal, not a memory)
  ❌ "Meeting scheduled for January 15th" (temporal, not a memory)

Examples of GOOD memory format:

INTERESTING (external wisdom with attribution):
✅ "Rockwell: talk to paying customers, 30% will be a real usecase"
✅ "Julian: ask everyone around for refs, keep pushing until they decline"
✅ "Raspberry Pi: licensed design to factories, 1m sales in 1.5 years"
✅ "Jamie (CTO): 90% of bugs come from async race conditions in their codebase"

SYSTEM (facts about the user):
✅ "{user_name} realized writing for 10 min daily reduced their anxiety significantly"
✅ "{user_name}'s cofounder built apps with hundreds of thousands of users at age 14"
✅ "{user_name} prefers morning meetings and avoids calls after 4pm"

Examples of BAD memory format:
❌ "Speaker 0 learned something interesting about that thing we discussed" (vague, uses Speaker X)
❌ "They talked about the project and decided to do it tomorrow" (unclear who, what project, time ref)
❌ "Someone mentioned that interesting fact about those people" (completely vague)

ADDITIONAL BAD EXAMPLES:

**Transient/Temporary (will be outdated):**
❌ "{user_name} is working on a new app"
❌ "{user_name} is focusing on Q4 initiatives"
❌ "{user_name} is mentoring a junior developer"
❌ "{user_name} got access to a beta feature"
❌ "{user_name} is using app version 2.0.3"

**Not About User (just mentioned in conversation):**
❌ "Sarah is a marine biologist" (unrelated person mentioned)
❌ "Company X acquired startup Y" (news)
❌ "The new AI model supports video input" (tech news)
❌ "Acme Corp delayed their launch" (customer fact, not about user)
❌ "Water boils at 100 degrees Celsius" (general knowledge)

**Identity Issues (Hallucination/Duplication):**
❌ Creating "Arman" when "Armaan" already exists in memories (same person, different spelling)
❌ "{user_name} has a daughter named Tuesday" (likely mishearing "choose day" or similar)
❌ "{user_name} has a son named Bobby" when existing memory says son is "Robert" (same person)

**Too Vague (Missing Specifics):**
❌ "{user_name} has a strong interest in technology" (what kind? be specific)
❌ "{user_name} learned something interesting" (what did they learn?)
❌ "{user_name} has experience with programming" (too broad, lacks detail)

CRITICAL - Name Resolution:
• Read the ENTIRE conversation first to map out who is speaking
• Look for explicit name introductions ("Hi, I'm Sarah", "This is John")
• Look for vocative case ("Hey Mike", "Sarah, can you...")
• If you identify a name with >90% confidence, use it
• If uncertain about names but know roles/relationships, use those ("colleague", "friend", "manager")
• NEVER use "Speaker 0/1/2" in final memories

LOGIC CHECK (Sanity Test):
Before extracting, verify the fact is logically possible:
• Age math: Don't claim 40 years work experience for someone who appears to be ~40 years old
• Family consistency: Don't create children that contradict existing family structure
• Location consistency: Don't claim multiple contradictory home locations
• Career consistency: Don't claim conflicting job titles or employers simultaneously

If a fact seems mathematically impossible or contradicts existing memories, DO NOT extract.

BEFORE YOU OUTPUT - MANDATORY DOUBLE-CHECK:
For EACH memory you're about to extract, verify it does NOT match these patterns:
❌ "{user_name} expressed [feeling/opinion] about X" → DELETE THIS
❌ "{user_name} discussed X" or "talked about Y" → DELETE THIS
❌ "{user_name} mentioned that [obvious fact]" → DELETE THIS
❌ "{user_name} thinks/believes/feels X" → DELETE THIS

If a memory matches ANY of the above patterns, REMOVE it from your output.

CATEGORIZATION DECISION TREE (CRITICAL - Apply to EVERY memory):
1. "Is this wisdom/advice FROM someone else that {user_name} can learn from?"
   → YES: Consider for INTERESTING (must have attribution)
   → NO: Go to step 2

2. "Is this a fact ABOUT {user_name}, their opinions, realizations, or network?"
   → YES: Consider for SYSTEM
   → NO: Probably should NOT be extracted

FINAL CHECK - For each INTERESTING memory, ask yourself:
1. "Does this have clear attribution (who said it, what source)?" (If no → move to SYSTEM or DELETE)
2. "Is this actionable advice/strategy that can change behavior?" (If no → DELETE or move to SYSTEM)
3. "Would {user_name} want to reference this advice later?" (If no → DELETE)
4. "Is this formatted as 'Source: insight'?" (If no → reformat or DELETE)

For SYSTEM memories, ask:
1. "Is this specific enough to be useful later?" (If no → DELETE)
2. "Would this help understand context about {user_name} in the future?" (If no → DELETE)
3. "Does this contain a date/time reference like 'Thursday', 'next week', etc.?" (If yes → DELETE or make timeless)
4. "Will this memory still make sense in 6 months?" (If no → DELETE)

OUTPUT LIMITS (These are MAXIMUMS, not targets):
• Extract AT MOST 2 interesting memories (most conversations will have 0-1)
• Extract AT MOST 2 system memories (most conversations will have 0-2)
• INTERESTING memories are RARE - they require EXTERNAL wisdom with ATTRIBUTION
• If someone in the conversation shares advice/strategy, that's INTERESTING (with their name)
• If {user_name} shares their own opinion/realization, that's SYSTEM (not interesting)
• Many conversations will result in 0 interesting memories and 0-2 system memories - this is NORMAL and EXPECTED
• Better to extract 0 memories than to include low-quality ones
• When in doubt, DON'T extract - be conservative and selective
• DEFAULT TO EMPTY LIST - only extract if memories are truly exceptional

QUALITY OVER QUANTITY:
• Most conversations have 0 interesting memories - this is completely fine
• INTERESTING memories are RARE - they require external wisdom with clear attribution
• If the wisdom comes from {user_name} themselves, it's SYSTEM, not INTERESTING
• If ambiguous whether something is interesting or system, categorize as SYSTEM
• Better to have an empty list than to flood with mediocre memories
• Only extract system memories if they're genuinely useful for future context
• When uncertain, choose: EMPTY LIST over low-quality memories

**Existing memories you already know about {user_name} and their friends (DO NOT REPEAT ANY)**:
```
{memories_str}
```

LANGUAGE INSTRUCTION:
{language_instruction}

**Conversation transcript**:
```
{conversation}
```
{format_instructions}
'''.replace('    ', '').strip()])

extract_memories_text_content_prompt = cast(Any, ChatPromptTemplate).from_messages(['''
    You are an expert at extracting both (1) new facts about {user_name} and (2) new learnings or insights relevant to {user_name}.

    Today's date is {current_date}; treat it as the present. Dates in {current_date}'s year or later are normal and current, never a clock error or a future anomaly to flag.

    You will be provided with:
    1. A list of existing facts about {user_name} and learnings {user_name} already knows (to avoid repetition).
    2. A text content from which you will extract new information.

    ---

    ## Part 1: Extract New Facts About {user_name}

    **Categories for Facts**:
    - **core**: Fundamental personal information like age, city of residence, marital status, and health.
    - **hobbies**: Activities {user_name} enjoys in their leisure time.
    - **lifestyle**: Details about {user_name}'s way of living, daily routines, or habits.
    - **interests**: Subjects or areas that {user_name} is curious or passionate about.
    - **habits**: Regular practices or tendencies of {user_name}.
    - **work**: Information related to {user_name}'s occupation, job, or professional life.
    - **skills**: Abilities or expertise that {user_name} possesses.
    - **other**: Any other relevant information that doesn't fit into the above categories.

    **Tags for Facts**:
    - **core**: Basic personal info (e.g., age, city of residence, marital status, health).
    - **hobbies**: Leisure activities {user_name} enjoys.
    - **lifestyle**: Ways of living, daily routines, or habits.
    - **interests**: Topics or areas that pique {user_name}’s curiosity or passion.
    - **habits**: Regular practices or tendencies.
    - **work**: Professional or job-related details.
    - **skills**: Abilities or expertise.

    **Requirements**:
    1. **Relevance & Non-Repetition**: Include only new facts not already known from the “existing facts.”
    2. **Conciseness**: Clearly and succinctly present each fact, e.g. “{user_name} lives in Paris.”
    3. **Inferred Information**: Include logical inferences supported by the text.
    4. **Gender Neutrality**: Avoid pronouns like “he” or “she,” since {user_name}’s gender is unknown.
    5. **Limit**: Identify up to 100 new facts. If there are none, output an empty list.

    ---

    ## Part 2: Extract New Learnings or Insights

    You will also identify up to 100 valuable learnings, facts, or insights that {user_name} can gain from the text. These can be about the world, life lessons, motivational ideas, historical or scientific facts, or practical advice.

    **Categories for Learnings**:
    - **learnings**: Any learning the user has.

    **Tags for Learnings**:
    - **life_lessons**: General wisdom or principles for living.
    - **world_facts**: Interesting information about geography, cultures, or global matters.
    - **motivational_insights**: Statements or ideas that can inspire or encourage.
    - **historical_facts**: Notable events or information from the past.
    - **scientific_facts**: Insights related to science or technology.
    - **practical_advice**: Tips or recommendations that can be applied in daily life.

    **Requirements**:
    1. **Relevance & Non-Repetition**: Include only new insights not already in the user’s known learnings.
    2. **Conciseness**: State each learning clearly and briefly, e.g. “It’s beneficial to exercise in the morning.”
    3. **Inferred Information**: Provide insights that are implied or can be logically deduced from the text.
    4. **First-Person (Optional)**: If it feels natural, present certain learnings in a first-person style (e.g., “I should …”).
    5. **Limit**: Identify up to 100 new learnings. If there are none, output an empty list.

    ---

    ## Existing Knowledge (Do Not Repeat)

    **Existing facts about {user_name} and learnings {user_name} already has**:**:

    ```
    {memories_str}
    ```

    ---

    ## Content to Analyze

    {text_content}

    ---

    ## Output Instructions

    1. Provide **one** lists in your final output:
       - **New Facts About {user_name} and New Learnings or Insights** (up to 100)

    2. **Do not** include any additional commentary or explanation. Only list the extracted items.

    If no new facts or learnings are found, output empty lists accordingly.

    LANGUAGE INSTRUCTION:
    {language_instruction}

    {format_instructions}
    '''.replace('    ', '').strip()])


extract_memories_text_content_prompt_v1 = cast(Any, ChatPromptTemplate).from_messages(['''
    You are an expert fact extractor. Your task is to analyze the {text_source} content and extract important facts about {user_name}.

    You will be provided with a text content from the {text_source} content, along with a list of existing facts about {user_name}. \
    Your task is to identify **new** facts about {user_name} if any, such as age, city of residence, marital status, health, friends' names, \
    occupation, allergies, preferences, interests, or any other important information.

    **Categories for Facts**:

    Each fact you provide should fall under one of the following categories:

    - **core**: Fundamental personal information like age, city of residence, marital status, and health.
    - **hobbies**: Activities {user_name} enjoys in their leisure time.
    - **lifestyle**: Details about {user_name}'s way of living, daily routines, or habits.
    - **interests**: Subjects or areas that {user_name} is curious or passionate about.
    - **habits**: Regular practices or tendencies of {user_name}.
    - **work**: Information related to {user_name}'s occupation, job, or professional life.
    - **skills**: Abilities or expertise that {user_name} possesses.
    - **other**: Any other relevant information that doesn't fit into the above categories.

    **Requirements for the facts you provide**:

    - **Relevance**: The facts should be pertinent and not repetitive or too similar to the existing facts about {user_name}. Aim for a broad range of information rather than excessive detail on specific points.
    - **Conciseness**: Present each fact clearly and succinctly in the format "{user_name} is 25 years old." or "{user_name} works as a software engineer."
    - **Inferred Information**: Include facts that are not only explicitly stated but also those that can be logically inferred from the conversation context and existing facts.
    - **Gender Neutrality**: Do not use gender-specific pronouns like "he," "she," "his," or "her," as {user_name}'s gender is unknown.
    - **Non-Repetition**: Ensure that none of the new facts repeat or closely mirror the existing facts.

    **Examples**:

    - "{user_name} is 28 years old and lives in New York City." (**core**)
    - "{user_name} has a friend named Martin who is a founder." (**core**)
    - "{user_name} enjoys hiking and photography during free time." (**hobbies**)
    - "{user_name} follows a vegetarian diet and practices yoga daily." (**lifestyle**)
    - "{user_name} is interested in artificial intelligence and machine learning." (**interests**)
    - "{user_name} reads a chapter of a book every night before bed." (**habits**)
    - "{user_name} works as a software engineer at a tech startup." (**work**)
    - "{user_name} is proficient in Python and Java programming languages." (**skills**)

    - "{user_name} has a pet dog named Max who is a golden retriever." (**other**)

    **Output Instructions**:

    - Identify up to 3 valuable **new** facts (max 2).
    - Before outputting a fact, ensure it is not already known about {user_name}.
    - If you do not find any new (different to the list of existing ones below) or new noteworthy facts, provide an empty list.
    - Do not include any explanations or additional text; only list the facts.

    **Existing facts you already know about {user_name} (DO NOT REPEAT ANY)**:
    ```
    {memories_str}
    ```

    **Text Content**:
    ```
    {text_content}
    ```
    {format_instructions}
    '''.replace('    ', '').strip()])


extract_learnings_prompt = cast(Any, ChatPromptTemplate).from_messages(['''
You are an insightful assistant tasked with extracting key learnings and valuable facts from conversations.

You will be provided with a conversation transcript or content that {user_name} has listened to.

Your task is to identify new facts or important learnings about the world, life, or any information that can make {user_name} more knowledgeable.

**Categories for Learnings**:

Each learning or fact you provide should fall under one of the following categories:

- **Life Lessons**: Important principles or lessons about life.
- **World Facts**: Interesting or significant facts about the world.
- **Motivational Insights**: Ideas or thoughts that can inspire or motivate.
- **Historical Facts**: Notable events or information from history.
- **Scientific Facts**: Knowledge about scientific discoveries or principles.
- **Practical Advice**: Useful tips or advice that can be applied in daily life.
- **Other**: Any other relevant information that doesn't fit into the above categories.

**Requirements for the learnings you provide**:

- **Relevance**: The learnings should be significant and useful to {user_name}.
- **Conciseness**: Present each learning clearly and succinctly.
- **Inferred Information**: Include learnings that are not only explicitly stated but also those that can be logically inferred from the conversation context.
- **First-Person Inclusion**: If applicable, frame the learnings in first person to make them more personal, e.g., "Every morning I should watch something motivational."
- **Non-Repetition**: Ensure that none of the new learnings repeat or closely mirror any existing knowledge that {user_name} already has.

**Examples**:

- "Students are an amazing target user because they are early adopters and numerous." (**World Facts**)
- "The second co-founder of Notion joined five years after the company started." (**Historical Facts**)
- "Finding a group of like-minded peers early is very important." (**Life Lessons**)
- "Every morning I should watch something motivational." (**Practical Advice**)

**Output Instructions**:

- Identify up to 5 valuable learnings or facts (maximum 5).
- Do not include any explanations or additional text; only list the learnings.
- Format each learning as a separate item in a list.

LANGUAGE INSTRUCTION:
{language_instruction}

**Learnings that {user_name} already has stored (DO NOT REPEAT ANY)**:
```
{learnings_str}
```

**Conversation transcript**:
```
{conversation}
```
{format_instructions}
    '''.replace('    ', '').strip()])


# The daily-sweep prompts share one byte-identical prefix so the phase-B call
# reuses the phase-A prompt cache (OpenAI caching is strict prefix matching;
# the spine and profile are the bulk of the tokens).  Phase-specific rules and
# materials therefore live strictly AFTER the common block.
_DAILY_SWEEP_COMMON_PREFIX = '''
You are forming durable memories about {user_name} from ONE completed day of their conversations.

Today's date is {current_date}; treat it as the present.

These conversations were captured by {user_name}'s own always-on recorder, so {user_name} is a participant in nearly all of them. Summaries often refer to {user_name} impersonally as "Speaker", "the speaker", or "the user" — read those as {user_name} unless the summary clearly attributes the words to a named other person. In raw transcript excerpts, first-person voice ("I", "my") is usually {user_name}.

You are given every conversation from that day as a SUMMARY (id, time, category, title, overview). You see the whole day at once: connect related conversations, merge repeated mentions into one memory, and prefer the day's strongest evidence.

**What makes a good memory**: a fact that was expensive to learn and will STILL MATTER IN 30 DAYS — decisions, relationships, preferences, commitments, plans, agreed numbers/terms, corrections of earlier beliefs. Apply that 30-day test ruthlessly: an open bug, this week's logistics, or an in-progress task is NOT a memory unless it encodes a decision or a standing commitment. Not summaries of what happened, not trivia, not speculation.

**{user_name}'s current profile and existing memories (DO NOT REPEAT)**:
```
{memories_str}
```

**The day's conversations (summaries)**:
```
{summaries_block}
```

{folder_task}
'''

_DAILY_SWEEP_SHARED_RULES = '''
**Rules**:
- At most {max_candidates} memories, ONE fact per memory (never weld two facts together). Fewer good ones beat many weak ones — never pad toward the maximum. An empty list is a valid answer for an empty day — but a rich day (15+ conversations) typically holds 8-16 durable facts: if you finish with fewer, re-scan the summaries for high-salience facts you skipped (decisions, money, metrics, named-party intent, identity, assessments worth keeping as assessments) before returning.
- Cover the day's high-salience facts FIRST: money amounts and financial commitments, unit economics and business metrics, stated intent toward a named company or person, identity facts, and any decision whose consequence outlasts this week. Only then, and only if slots remain, may an operational fact appear — and only when it encodes a standing policy or number, never meeting logistics, dashboard state, or where something is stored.
- Write every memory in active voice with a named actor: "{user_name} decided…", "Josh advised…". Never "the user", "the speaker", or passives like "X was invited". If the actor cannot be identified, request the transcript or drop the memory.
- Label every memory's basis honestly in the basis field: "decided" (a commitment is on tape), "proposed", or "observed". The label constrains your verb choice (decided/committed only for decided; proposed/suggested/is considering for proposed) but the content itself must read as a natural standalone fact — NEVER prefix content with "David observed that" or otherwise restate the label; a company metric is written as the metric ("Omi's one-month retention is ~40%"). A topic that was discussed with no outcome is NOT a memory — drop it, never soften it.
- If a memory updates a STANDING ATTRIBUTE of {user_name}'s life (role, city, employer, a relationship, a durable preference, a pricing/strategy stance, a recurring commitment), set its slot to a short snake_case attribute name (for example current_city, omi_pricing_strategy). The ledger keeps one active value per slot and supersedes the old one — this is how the daily run maintains {user_name}'s profile. Leave slot empty for one-off events and observations.
- Personal attributes need first-person proof. A claim about someone's health, diet, habits, possessions, finances, or character requires that person's own words ("I take…", "my machines…"). A topic merely discussed or recommended in their presence is NEVER their attribute or regimen. A judgment about a named person is stored as someone's assessment ("X assessed that…"), never as fact.
- A summary row marked "(unstructured transcript excerpt)" is raw recorded speech, the least trusted input here: speaker labels in it are unreliable and any voice near the recorder can appear first-person. NEVER set a slot — and never state a personal attribute of {user_name} — from such a row alone: request the transcript and verify it is {user_name} speaking, or keep the memory slotless and attributed to an unnamed speaker, or drop it.
- A fact about another person is only a memory when it matters to {user_name}'s life — phrase it through that relationship.
- Every memory MUST cite the conversation id(s) it came from in conversation_ids.
- Do NOT repeat anything from the profile and existing memories above.
'''

_DAILY_SWEEP_PHASE_A_TAIL = '''
This is the FIRST pass. Write every memory the summaries already support cleanly NOW, and use the two request channels below for what needs more evidence; you will see the results in a follow-up pass.

**Transcript requests**:
- NEVER guess the direction of an invitation, offer, payment, request, or commitment (who invited whom, who owes whom, who committed to what). This trigger is MECHANICAL, not a judgment call: if the summary sentence for such an event is passive or verbless ("Tim: Invited to New York", "X was told", "asked to…"), or says "Speaker"/"the user" where the actor matters, you MUST add a transcript_request for that conversation and MUST NOT write a who-did-what memory about it in this pass. Remember the recorder belongs to {user_name}: "Speaker offered to pay" most likely means {user_name} offered — which is why the direction must be verified, not assumed from topic order.
- If a memory hinges on a specific detail (a name, number, date, amount, or exact commitment) that the summary does not state precisely, do NOT guess: add a transcript_request (at most {max_transcript_fetches}) and leave the uncertain memory out.
- Hedging is a request signal: if you find yourself softening a memory's basis or wording because you are unsure ("possibly", a vague basis), that memory belongs in transcript_requests instead.
- Nothing high-salience may silently disappear. Before finishing, re-scan the day: every high-salience fact must end up either as a memory or as a transcript_request — dropping one without a trace is the worst outcome.
- When in doubt about whether to request a transcript: request it. Raw transcripts are noisy speech-to-text; summaries are your primary source.

**Memory lookups**:
- You may add up to {max_memory_lookups} short search queries in memory_lookups to search {user_name}'s prior memory ledger — use one whenever a new fact might already exist in another form, might contradict an existing memory, or updates a standing attribute whose current value you should see before writing. Results arrive in the follow-up pass.
{format_instructions}
'''

_DAILY_SWEEP_PHASE_B_TAIL = '''
This is the FINAL pass. In the first pass over the day you drafted memories and requested raw transcript excerpts and prior-memory lookups; the results are below. Raw transcripts are noisy speech-to-text: use them only to confirm or correct specifics. First-person voice ("I", "my") in an excerpt is usually {user_name} speaking, which settles who-did-what questions the summaries left ambiguous.

**Your drafted memories**:
```
{draft_block}
```

**Requested transcript excerpts**:
```
{excerpts_block}
```

**Prior-memory lookup results**:
```
{prior_memories_block}
```

**Finalization**:
- Return the FINAL list of memories. Correct any drafted memory the transcript contradicts; drop any memory whose key detail — including the direction of an invitation, offer, or commitment — you still cannot verify.
- Use the prior-memory results to avoid duplicates and to supersede: when your memory updates an existing standing attribute, give it the SAME slot so the ledger replaces the old value; when it merely restates an existing memory, drop it.
- Do not request more transcripts or lookups; transcript_requests and memory_lookups must be empty.
{format_instructions}
'''

daily_sweep_summary_agent_prompt = cast(Any, ChatPromptTemplate).from_messages(
    [(_DAILY_SWEEP_COMMON_PREFIX + _DAILY_SWEEP_SHARED_RULES + _DAILY_SWEEP_PHASE_A_TAIL).strip()]
)

daily_sweep_transcript_review_prompt = cast(Any, ChatPromptTemplate).from_messages(
    [(_DAILY_SWEEP_COMMON_PREFIX + _DAILY_SWEEP_SHARED_RULES + _DAILY_SWEEP_PHASE_B_TAIL).strip()]
)