Agentic Design

Patterns
🧠

Parametric Memory(PM)

Knowledge implicitly stored within model parameters, enabling fast context-free knowledge retrieval for multi-agent agentic AI systems

Complexity: mediumMemory Management

🎯 30-Second Overview

Pattern: Knowledge implicitly stored within model parameters, enabling fast context-free access

Why: Sub-millisecond retrieval, no external dependencies, consistent across agents, scales to 1000+ agents

Key Insight: Pre-trained knowledge → PEFT specialization → shared base parameters across agent network

⚡ Quick Implementation

1Pre-train:Embed domain knowledge in model weights
2Fine-tune:Use PEFT methods (LoRA, QLoRA) for efficiency
3Specialize:Create agent-specific parameter branches
4Share Base:Common foundation across agent network
5Monitor:Track knowledge access & parameter usage
Example: base_llm → domain_finetune → agent_specialization → shared_deployment

📋 Do's & Don'ts

✅Use parameter-efficient fine-tuning (LoRA, QLoRA, MoRA)
✅Share base parameters across agent network for consistency
✅Monitor parameter specialization patterns for knowledge storage
✅Implement knowledge consolidation to prevent parameter bloat
✅Version control model parameters for rollback capability
❌Full parameter retraining for new knowledge (expensive & risky)
❌Store time-sensitive information in parametric memory
❌Ignore knowledge cutoff dates and factual accuracy degradation
❌Deploy without parameter redundancy for fault tolerance
❌Mix incompatible parameter versions across agent instances

🚦 When to Use

Use When

  • • Stable domain knowledge required
  • • Fast inference speed critical
  • • Multi-agent consistency needed
  • • Offline deployment scenarios
  • • Cost-sensitive applications

Avoid When

  • • Rapidly changing information
  • • Regulatory compliance updates
  • • Real-time data integration
  • • User-specific customization
  • • Frequent knowledge updates

📊 Key Metrics

Knowledge Accuracy
% correct factual responses
Inference Speed
Tokens/second generation rate
Parameter Efficiency
Knowledge/parameter ratio
Consistency Score
Cross-agent response similarity
Memory Footprint
GB required per agent
Update Cost
$ per knowledge refresh cycle

💡 Top Use Cases

Scientific Research Agents: Physics, chemistry, biology knowledge embedded (consistent across 100+ agents)
Legal Document Analysis: Case law, statutes, procedures in parameters (instant access, no external DB)
Medical Diagnosis Support: Medical knowledge, drug interactions, symptoms (HIPAA-compliant, offline)
Financial Analysis: Market fundamentals, accounting principles, regulations (real-time inference)
Code Generation: Programming languages, frameworks, best practices (multi-language consistency)

References & Further Reading

Deepen your understanding with these curated resources

Academic Papers

A Survey on the Memory Mechanism of Large Language Model based Agents (Zhang et al., 2024)

The Rise of Parameter Specialization for Knowledge Storage (Hong et al., 2024)

Parametric vs Non-parametric Memory in Retrieval-augmented LLMs (Farahani & Johansson, 2024)

Augmented Large Language Models with Parametric Knowledge Guiding (Luo et al., 2023)

Parameter-Efficient Methods

MoDE: Multi-task Parameter Efficient Fine-Tuning (Ning et al., 2024)

L4Q: Parameter Efficient Quantization-Aware Fine-Tuning (Jeon et al., 2024)

UNLEARN: Efficient Knowledge Removal in LLMs (Lizzo & Heck, 2024)

LoRA: Low-Rank Adaptation of Large Language Models (Hu et al., 2021)

Multi-Agent Systems

Multi-Agent Collaboration Mechanisms: A Survey of LLMs (2024)

LLM Multi-Agent Systems: Challenges and Open Problems (2024)

LLM Agent Memory Survey - GitHub Repository

Efficient LLMs Survey - GitHub Repository

Industry Implementation

Anthropic Claude Memory Management Documentation

Model Context Protocol (MCP) - Anthropic Standard

Hugging Face PEFT (Parameter-Efficient Fine-Tuning) Library

OpenAI Fine-tuning Cookbook - How to Fine-tune Chat Models

Contribute to this collection

Know a great resource? Submit a pull request to add it.

Contribute

Patterns

closed

Loading...

Built by Kortexya