About this ebook
Storing a conversation is not memory. Retrieving an old document is not memory. Even perfectly preserving everything that ever happened does not guarantee that an AI will make a better decision today.
Memory starts when the past changes present behaviour — and changes it for the better.
In Memory, Ernan Hughes builds that idea from first principles. Rather than beginning with a taxonomy of human memory or assuming that more sophisticated architectures must be better, the book starts with simple systems, measures where they fail, and adds complexity only when the evidence demands it.
The investigation begins with six practical questions:
Where did we discuss this?
What did we decide?
Why did we decide it?
Is it still true?
What did we leave unfinished?
What from the past matters right now?
From those questions emerges a practical architecture for AI memory.
The book examines strong RAG, persistent structured memory, associative retrieval, provenance, temporal state, unfinished work, context selection, context assembly, trust, long-term growth, and the boundary between memory and learning. Each mechanism has to justify itself against simpler alternatives.
Along the way, experiments reveal several important results.
More history can make an AI perform worse than having no memory at all. Strong retrieval remains surprisingly difficult to beat. Persistent interpretations can preserve useful understanding — but they can also preserve mistakes. A memory can be historically true while no longer being safe to act upon. Selecting the right memories and assembling them into a bounded context are separate problems. And once remembered information can influence action, memory becomes an authority and security problem as well as a retrieval problem.
The result is not an architecture designed to imitate the human brain.
It is an engineering framework for deciding what from the past should be allowed to influence what happens now.
Throughout the book, implementations, controlled experiments, negative results, ablations, notebooks, and a final capstone system are used to test the ideas rather than merely describe them.
You will learn how to:
distinguish storage, retrieval, context, memory, and learning;
build serious retrieval baselines before adding complexity;
preserve decisions, provenance, corrections, and supersession over time;
represent unfinished work and current project state;
select useful memory under a limited context budget;
measure whether remembered information actually changes behaviour;
detect when memory makes an AI worse rather than better;
keep derived memories auditable, reversible, and traceable to their evidence;
design memory systems that can evolve without silently rewriting their own history.
Memory is for engineers, researchers, and practitioners building AI agents, assistants, coding systems, and other applications that must operate across sessions, projects, and long-running histories.
The goal is not to make AI remember everything.
The goal is to make the present better because the past happened.