---

name: knowledge-extraction description: Use when processing conversations to extract reusable knowledge (entities, facts, processes) trigger_pattern: When ingesting a conversation, processing user input, or consolidating memories tags:

• meta
• extraction
• entity
• fact
• process version: 1 is_protected: true q_value: 0.5 q_update_count: 0

---

Knowledge Extraction

Iron Law

EXTRACT ONLY WHAT'S WORTH REMEMBERING FOR FUTURE CONVERSATIONS.

Not everything in a conversation is knowledge. Filter aggressively.

Process Extraction Principle: Quality over Quantity

Better to extract 1 high-quality, complete Process than 10 sloppy ones.

Each Process should be like a "reviewable story":

• A newcomer can understand the complete context after seeing it
• They can judge whether it applies to the current similar scenario
• They know why this method works

When to Extract

Trigger Conditions:
├── New conversation received for memory storage
├── User explicitly shares a preference or fact
├── A successful procedure/workflow is completed
├── User corrects previous understanding
└── Pattern emerges across multiple interactions

What to Extract

1. Entities

Real-world things that have attributes worth tracking.

Types:

• PERSON: Users, team members, contacts
• PROJECT: Codebases, products, initiatives
• ORGANIZATION: Companies, teams, departments
• TOOL: Software, libraries, frameworks
• CONCEPT: Domain concepts, technical terms

Extract When:

• Entity has attributes mentioned in conversation
• Entity is referenced multiple times
• Entity relationships are established

DON'T Extract:

• Random variable names from code
• Temporary values or config entries
• Error message fragments
• Generic terms without context

2. Facts (Attributes)

Properties of entities with scope and cardinality.

Structure:

Entity -[HAS_ATTRIBUTE]-> Fact
  - slot: "preference.theme" | "tech.database" | "role.title"
  - value: "dark" | "PostgreSQL" | "Tech Lead"
  - cardinality: "single" | "multi"
  - scope: "universal" | "project" | "task" | "session"

Cardinality Rules:

Cardinality	Meaning	Conflict Handling	Examples
single	Only one value valid	New value supersedes old	job title, preferred theme
multi	Multiple values coexist	Deduplicate only	hobbies, skills, languages

Scope Rules:

Scope	Meaning	Persistence	Examples
universal	Always true	Permanent	"I prefer dark mode"
project	True within project	Project lifetime	"This project uses PostgreSQL"
task	True for current task	Task lifetime	"Currently debugging OOM"
session	Temporary choice	DON'T PERSIST	"Let's try this approach first"

Extract When:

• User EXPLICITLY states preference: "I like...", "I prefer...", "I always..."
• User CONFIRMS a fact: "Yes, we use X", "That's correct"
• Learned configuration that user confirmed works

DON'T Extract:

• Assistant suggestions not confirmed by user
• Implementation details (passwords, ports, config values)
• Debugging artifacts (error messages, stack traces)
• Session-scoped temporary decisions

3. Processes

Trigger → Action → Outcome patterns that may become skills.

Structure:

Process {
  trigger: "When/If [condition]"      # Complete situation description
  action: "Do [steps]"
  outcome: "Result [expected state]"
  context: "Project/tech stack/constraint background"       # NEW
  problem_statement: "Specific problem to solve"  # NEW
  key_insight: "Why this method works"      # NEW
  is_generalizable: bool
}

Trigger Quality Requirements:

Trigger is not a simple conditional statement, but a complete situation description.

Comparison	Example
BAD	"When writing Go HTTP handlers"
GOOD	"When implementing a new REST endpoint in a Go project (Gin framework, clean architecture), needing JSON validation, error handling, Swagger documentation, unit tests, and the team wants to establish consistent handler patterns"

Trigger must include:

1. Task: What you are doing (specific task)
2. Context: Project/tech stack/team background
3. Problem/Need: Why you need this Process (pain point or goal)

New Field Extraction Guidance:

Field	Source	Example
context	Project name, framework, team conventions mentioned in conversation	"h-mem project, using Neo4j + Python"
problem_statement	Pain point or requirement described by user	"Handler style is inconsistent, need to establish a pattern"
key_insight	Why the solution works	"DTO-first design ensures type safety"

Generalizability Test:

Is this process GENERALIZABLE?
├── Can it apply to different projects?
│   └── YES → is_generalizable = true
├── Can it apply to different contexts?
│   └── YES → is_generalizable = true
├── Would another agent benefit from knowing this?
│   └── YES → is_generalizable = true
├── Is it a one-time specific fix?
│   └── NO → is_generalizable = false
└── Is it tied to specific values/passwords?
    └── NO → is_generalizable = false

Generalizable Examples ✓:

• "When OOM occurs → Capture heap dump → Identify leak source"
• "Before deployment → Run full test suite → Ensure stability"
• "When designing API → Define interfaces first → Consistent contracts"

NOT Generalizable ✗:

• "Set password to testpassword123" (specific value)
• "Restart the Neo4j container" (specific debugging step)
• "Change line 234 in auth.py" (too specific)
• "Run npm install" (trivial command)

Extraction Quality Checklist

Before storing extracted knowledge:

• Uniqueness: Is this information already stored?
• Relevance: Will this help in future conversations?
• Accuracy: Did the user confirm this, or is it assumed?
• Scope: Is the scope correctly classified?
• Cardinality: Single or multi? Will conflict detection work?
• Generalizability: For processes, can it transfer to new contexts?
• No Secrets: No passwords, API keys, or sensitive data?

Anti-patterns

Anti-pattern	Why It's Wrong	Correct Approach
Extract everything	Pollutes memory with noise	Filter aggressively
Assume user intent	Creates false knowledge	Only extract explicit statements
Store session decisions	Clutters with temporary info	Mark as session scope, don't persist
Store debugging steps	Not reusable	Only store generalizable processes
Duplicate information	Wastes space, causes conflicts	Deduplicate by (entity, slot, value)
Store assistant suggestions	User didn't confirm	Wait for user confirmation

Extraction Pipeline

Conversation Received
    │
    ▼
┌─────────────────────────────────────────┐
│ 1. Read ENTIRE conversation first       │
│    (Don't extract per-message)          │
└─────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────┐
│ 2. Identify Entities with attributes    │
│    - Skip entities without facts        │
└─────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────┐
│ 3. Extract Facts with scope + cardinality│
│    - Skip session-scoped (don't persist)│
│    - Deduplicate by (entity, slot, value)│
└─────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────┐
│ 4. Identify Processes                   │
│    - Check generalizability             │
│    - Skip non-generalizable             │
└─────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────┐
│ 5. Store in Neo4j with provenance       │
│    Conversation -[:GENERATES]-> Fact    │
│    Conversation -[:GENERATES]-> Process │
│    Entity -[:HAS_ATTRIBUTE]-> Fact      │
└─────────────────────────────────────────┘

Conflict Resolution

When a new fact conflicts with existing knowledge:

Scenario	Resolution
Same entity, same slot, different value, cardinality=single	New supersedes old (mark old as superseded)
Same entity, same slot, same value	Deduplicate (don't create new)
Same entity, same slot, different value, cardinality=multi	Add new value (coexist)
Different scope	Both can coexist (universal ≠ project-specific)

Examples

Good Extraction

Conversation:

User: I always use dark mode, it's easier on my eyes.
Assistant: I'll remember that preference.
User: Also, for this h-mem project, we're using Neo4j as the database.

Extracted:

entities:
  - name: "User"
    type: "PERSON"
  - name: "h-mem"
    type: "PROJECT"

facts:
  - entity: "User"
    slot: "preference.theme"
    value: "dark"
    cardinality: "single"
    scope: "universal"  # "always" = universal

  - entity: "h-mem"
    slot: "tech.database"
    value: "Neo4j"
    cardinality: "single"
    scope: "project"  # specific to h-mem

processes: []  # No generalizable procedures here

Process Extraction: Good vs Bad

Conversation:

User: We need to add a new REST endpoint for user registration in our Go project.
      We're using Gin framework with clean architecture. I want proper validation,
      error handling, and Swagger docs. We've had issues with inconsistent handler
      patterns across the team.
Assistant: [Implements the endpoint with DTO-first approach, validation middleware,
           structured error responses, and Swagger annotations]
User: This is exactly what I wanted. The DTO approach really helps keep things organized.

BAD Process Extraction (Too Simplified):

processes:
  - trigger: "When writing Go HTTP handlers"
    action: "Use DTOs and middleware"
    outcome: "Handler is created"
    is_generalizable: true
    # Missing: context, problem_statement, key_insight
    # Trigger too vague - doesn't capture the full situation

GOOD Process Extraction (Complete Story):

processes:
  - trigger: "When implementing a new REST endpoint in a Go project (Gin framework, clean architecture), needing JSON validation, error handling, Swagger documentation, and the team has handler style inconsistency issues"
    action: |
      1. Define Request/Response DTO structs (with binding tags)
      2. Create validation middleware to handle validation errors uniformly
      3. Use structured error response format
      4. Add Swagger annotations to generate documentation
      5. Handler only does DTO conversion and calls the service layer
    outcome: "Handler follows a consistent pattern that the team can replicate to other endpoints"
    context: "Go + Gin + Clean Architecture project"
    problem_statement: "Team handler style is inconsistent, need to establish a repeatable pattern"
    key_insight: "DTO-first design separates validation logic from business logic, ensuring type safety and testability"
    is_generalizable: true

Bad Extraction (Don't Do This)

Same Conversation, Wrong Extraction:

# WRONG - extracted too much
entities:
  - name: "dark mode"      # Not an entity, it's a value
  - name: "eyes"           # Irrelevant
  - name: "Neo4j"          # Should be a value, not entity

facts:
  - entity: "User"
    slot: "reason"
    value: "easier on eyes"  # Explanation, not a fact
    scope: "session"         # Wrong scope