Foundations

AI Memory

How AI systems remember across turns, sessions, and lifetimes. The base layer is vector-based retrieval. On top of that sit the established options (Mem0, Zep, Letta), a wave of new research, and the memory features built into ChatGPT, Claude, and Gemini. This page covers what is real, what is emerging, and where the field is heading.

In this topic

Everything about AI memory

Reference

A complete walk through AI memory: why stateless models make memory the central problem, vector-based retrieval as the foundation, the established frameworks (Mem0, Zep/Graphiti, Letta/MemGPT) compared, frontier 2026 research (MAGMA, ZenBrain, field-theoretic memory, Second Me, GSW, OCR-Memory, MemMachine), the survey taxonomy (factual / experiential / working memory and formation / evolution / retrieval dynamics), raw vs derived as the central design spectrum, LeCun's world models and the A-B-M framework with System M as the meta-controller, open problems, vendor memory products (Microsoft GitHub Copilot Memory, OpenAI ChatGPT Memory, Google Gemini Memory Bank, Anthropic Auto Dream, AWS AgentCore Memory + Bedrock Knowledge Bases), and how memory connects to agents.

Last updated May 16, 2026