All terms
Agents & tools
AI memory
Also known as: persistent memory, long-term memory, cross-chat memory
Persistent context an AI assistant carries across separate conversations — your preferences, past projects, recurring facts — so you don't have to re-explain yourself every time.
What it means
AI memory is the catch-all term for "the assistant remembers things between chats." ChatGPT Memory, Claude's project-level memory, and various third-party layers (Mem, Letta, Zep) all promise the same thing: you tell it once, it remembers forever.
What's actually happening is almost never magic. Under the hood, "memory" is a retrieval system: the assistant writes down salient facts during your conversations (often via a separate "memory updater" LLM call), stores them in a database, and at the start of each new conversation retrieves the relevant ones and stuffs them into the system prompt. It's RAG over a personal knowledge base. The model doesn't actually "remember" — its weights don't change — but from the user's perspective it acts like it does.
This means memory has all the failure modes of retrieval: it can pull stale facts, miss relevant ones, or pollute context with irrelevant trivia. Worse, since memory writes are also LLM calls, the system can confidently misremember things you never said. ChatGPT users have hit cases where memory contains wrong inferences ("user is allergic to peanuts" when you only mentioned a friend was). Most products give you a memory inspector for this reason — use it.
By 2026 the landscape is roughly: ChatGPT Memory (cross-chat, opt-in, edit anytime), Claude Memory (project-scoped or user-scoped depending on tier), Cursor and Claude Code (project-local CLAUDE.md / .cursorrules files — manually written, not learned), and a wave of agent frameworks where you build memory yourself with vector stores. Real long-term memory — model weights that update from your interactions — remains a research problem.
Example
You tell ChatGPT "I prefer Python over JavaScript and I'm vegetarian." Three weeks later in a totally separate chat, you ask for a meal plan and it suggests vegetarian recipes — because at the start of that chat it retrieved both facts into context.
Why it matters
Memory is what turns an AI tool into a relationship. It's also where users notice the difference between assistants the most. But it's mostly retrieval under a marketing layer — understanding that helps you debug it ("why did it forget?" usually means a retrieval miss) and design better systems if you're building one.