Memory System Design

Design memory architectures for agents with layered persistence from working memory to knowledge graphs.

Prerequisites

•Understanding of vector embeddings
•Familiarity with graph databases

Instructions

Memory Layers

Layer	Persistence	Access	Use Case
Working	Context window	Zero latency	Current task
Short-term	Session	Searchable	Conversation state
Long-term	Cross-session	Structured	Learning over time
Entity	Permanent	Queryable	Entity consistency

Memory Architecture Comparison

System	Accuracy	Latency	Notes
Temporal KG	94.8%	2.58s	Best accuracy
MemGPT	93.4%	Variable	Good general
GraphRAG	75-85%	Variable	20-35% over RAG
Vector RAG	60-70%	Fast	Loses relationships

Why Vector Stores Fall Short

Vector stores lose relationship information. If agent learns "Customer X purchased Product Y on Date Z," it cannot answer "What did customers who bought Y also buy?"

Knowledge Graph Advantages

•Preserve relationships between entities
•Enable queries that traverse relationships
•Support temporal validity periods

Temporal Knowledge Graphs

Add validity periods to facts:

python

# What was user's address on January 15?
temporal_graph.query("""
    MATCH (user)-[r:LIVES_AT]->(address)
    WHERE r.valid_from <= $date AND r.valid_until > $date
    RETURN address
""", {"date": "2024-01-15"})

Memory Selection Guide

•Simple persistence: File-system memory
•Semantic search: Vector RAG with metadata
•Relationship reasoning: Knowledge graph
•Temporal validity: Temporal knowledge graph

Guidelines

•Match architecture to query requirements
•Use temporal validity to prevent outdated conflicts
•Consolidate memories periodically
•Design for retrieval failures gracefully
•Consider privacy implications of persistence

Notes

•Zep showed 90% retrieval latency reduction vs full-context
•GraphRAG reduces hallucination by up to 30%
•Entity memory maintains consistency across conversations

Source: muratcankoylan/Agent-Skills-for-Context-Engineering