RAG vs knowledge graph vs memory: which does a personal AI need?

Retrieval-augmented generation lets a language model fetch information it was not trained on, then answer from it; the name comes from a 2020 paper. It finds what matches the words of a question: last year's price sheet, an old thread, an invoice.

IBM's comparison puts it plainly: retrieval "is like giving the home cook a cookbook for that cuisine." And keeping the cookbook current is the cheap part: "it's cheaper to keep retrieval indices up-to-date than to continuously pre-train an LLM," as Eugene Yan writes.

A knowledge graph stores things and the relations between them: in the encyclopedia's words, "a knowledge base that uses a graph-structured data model or topology to represent and operate on data." For a buyer asking for a discount, it connects the buyer, the account, the past orders and the past discounts.

It answers "what relates to what". It does not say which of those facts should decide this request.

Memory keeps what you have told an assistant, and lets a newer fact replace an older one. One memory vendor's documentation puts the difference in a line: "RAG answers 'What do I know?' while Memory answers 'What do I remember about you?'"

Its example is a shopper who loves Adidas on day 1 and switches to Puma on day 31. Asked on day 45 what to buy, retrieval can still surface the old praise; memory knows it is out of date.

None of the three says what you would do with a request. Your engram stores that: your decisions, your rules and your way of saying things, each traced to where you said it, with the dates it applies, so a reversed rule replaces the old one the way memory replaces an old preference.

It is built on the other three, not instead of them, and it is measured against you: decisions you already made are held back and replayed, and the score is how often it would have done the same.