Comparing Memory Systems for LLM Agents: Vector, Graph, and Event Logs

Abstraction: Agent memory patterns compared by latency, hit-rate, failure modes

Key points:

• Compares 6 memory patterns in 3 families: vector (plain RAG, MemGPT-style tiered), graph (Zep/Graphiti temporal KG, Microsoft GraphRAG), and event/logs (ALAS, LangGraph checkpoints, episodic memory).
• Vector RAG (FAISS/HNSW/ScaNN) gives sublinear near-logarithmic ANN latency, low tens of ms for millions of items, but fails on temporal, cross-session, and multi-hop queries — motivating benchmarks DMR and LongMemEval.
• Zep/Graphiti temporal KG reports 94.8% vs 93.4% on DMR over MemGPT, and up to 18.5% higher accuracy plus ~90% lower latency than baselines on LongMemEval.
• Execution logs (ALAS with validator isolation + localized cascading repair; LangGraph thread-scoped checkpoints) give ~100% hit rate for "what happened" queries, enabling replay/repair, but risk log bloat, partial instrumentation, and unsafe side-effect replay.
• Key thesis: memory is a systems problem; robust agents compose vector + graph + event/episodic layers rather than one "magic" mechanism.

Connections: Memgpt · Zep · Graphrag · Langgraph · AI Agents · Retrieval Augmented Generation · Knowledge Graphs · Agent Memory

Source: https://www.marktechpost.com/2025/11/10/comparing-memory-systems-for-llm-agents-vector-graph-and-event-logs/