Knowledge Representation

Knowledge Representation

type: utility mode: generative triggers:

• "knowledge management"
• "context preservation"
• "agent memory"
• "knowledge schema"
• "information retrieval"

Purpose

Design effective knowledge representation and retrieval systems for multi-agent and single-agent systems. Provides frameworks for organizing information, managing context decay, and enabling efficient retrieval.

Core Principle

Remember what matters, find it when needed. The goal isn't to remember everything—it's to identify what matters, encode relationships that enable reasoning, and retrieve the right information at the right time.

Knowledge Hierarchy (DIKW)

Context Decay Patterns

Knowledge relevance decreases over time:

Layered Context Model

Organize context by stability and scope:

Global Context (Always Relevant)

• Project goals and constraints
• Architectural principles
• Security requirements
• Performance targets
• Size limit: ~1000 tokens
• Update frequency: Rare

Session Context (Current Work)

• Active task details
• Recent decisions
• Working assumptions
• Open questions
• Size limit: ~5000 tokens
• Update frequency: Per task

Local Context (Current Operation)

• Immediate file contents
• Current function scope
• Active variables
• Error states
• Size limit: ~10000 tokens
• Update frequency: Continuous

Context Compression Techniques

Summarization

• Progressive abstraction
• Compression ratio: 5:1
• Information loss: Moderate

Reference Substitution

• Replace details with pointers: See UserService.ts:L45-L550
• Compression ratio: 100:1
• Information loss: None (if accessible)

Pattern Extraction

• Identify and reference patterns
• Example: "Standard error pattern x10"
• Compression ratio: 10:1

Retrieval Strategies

Vector Similarity Search

1. Generate query embedding
2. Compare with stored embeddings
3. Rank by similarity score
4. Apply threshold filter
5. Return top-k results

Faceted Search

Filter by multiple dimensions:

• Domain (frontend, backend, etc.)
• Complexity (trivial, small, medium, large)
• Tags (bug, feature, refactor)
• Date range

Query Expansion

Memory Types

Working Memory

• Capacity: 7 ± 2 items
• Focus item with full attention
• Context items with decreasing weights
• LRU replacement policy

Episodic Memory

• Records of specific experiences
• What happened, when, what was learned
• Retrieval cues: similar errors, same location

Semantic Memory

• General knowledge about concepts
• Relationships between concepts
• Common issues and solutions

Procedural Memory

• Knowledge of how to perform tasks
• Step-by-step processes
• Tools and patterns for each step

Quality Management

Confidence Scoring

Age Decay Formula

confidence = base_confidence × (0.5 ^ (age_days / half_life))

Contradiction Detection

• Direct contradictions (A vs not-A)
• Temporal contradictions (timeline impossibilities)
• Semantic contradictions (conceptual conflicts)

Anti-Patterns

"Everything is Important"

Problem: 100,000+ items, 15% precision, 1:20 signal-to-noise

Solution: 5,000 curated items, 85% precision, 15:1 signal-to-noise

"Never Forget"

Problem: Never delete, 1GB/month growth, 85% stale

Solution: Managed lifecycle with retention policies:

• Critical: Permanent
• Important: 1 year
• Operational: 3 months
• Temporary: 1 week

"Context Bloat"

Problem: 50,000 tokens, 500 relevant, 30 sec processing

Solution: 5,000 tokens, 4,500 relevant, 2 sec processing

Implementation Checklist

• Define knowledge hierarchy levels
• Establish decay policies by knowledge type
• Implement layered context model
• Set up compression strategies
• Configure retrieval mechanisms
• Establish confidence scoring
• Implement contradiction detection
• Create knowledge validation processes

Integration Points

Inbound:

• When designing agent systems
• When planning knowledge storage
• When context limits become a problem

Outbound:

• To agent architecture design
• To context network development
• To retrieval system implementation

Complementary:

• context-retrospective: For evaluating knowledge system effectiveness
• task-decomposition: For understanding task knowledge needs