What Remembering Means
Storage, retrieval, and memory are three different things.
Read chapter →
Programmer.ie Book
In development
Memory
How AI systems turn past experience into better present behaviour.
Storage, retrieval, and memory are different things.
The central idea of this book is simple: memory is when past experience changes present behaviour.
Rather than beginning with a taxonomy of working, episodic, semantic, procedural, entity, and task memory, the book begins with six questions a useful project memory should be able to answer:
- Where did we discuss X?
- What did we decide about X?
- Why did we decide it?
- Is it still true?
- What did we leave unfinished?
- What should I remember right now?
Each step follows the same method: build the strongest simple mechanism first, measure it, diagnose the failure, add only the mechanism the evidence earns, and rerun the same harness.
The investigation now closes in three stages. Chapter 15 asks how a long-lived memory controls growth without destroying historical recoverability. Chapter 16 asks where remembering ends and learning from outcomes begins. Chapter 17 reconstructs the smallest architecture the experiments actually earned.
Chapter 12 is the behavioural hinge: it tests whether retained project history changes present action under matched conditions. Its result is deliberately not assumed by the final chapters until a canonical run exists.
Chapters
The Measurement Instrument
Define, validate, and build the instrument we will use to measure whether an AI system actually remembers.
Read chapter →The RAG Baseline
Build conventional retrieval-augmented memory, inspect what reaches the reader, and establish the baseline every later mechanism must beat.
Read chapter →From Retrieval to Persistent Understanding
Strong RAG already understands retrieved evidence at query time. This chapter asks what changes when some of that understanding is preserved as reusable state, builds a persistent derived graph, and measures what it costs.
Read chapter →Pathways Through Memory
Explore how activation can propagate through a memory graph so that one remembered thing leads to another, and measure whether associative retrieval improves on static graph search and conventional RAG.
Read chapter →The Memory Nexus
Once a system has several ways to remember, something has to choose between them. This chapter builds that control layer, measures it against simpler alternatives, and lets the evidence decide whether it earns its place.
Read chapter →Why
Trace beliefs and generated claims through derived memory back to the evidence that actually supports them, distinguishing support from retrieval, derivation, repetition, and mere citation.
Read chapter →Is It Still True?
Some information lives in the ordered transitions between remembered states. A ZeroMQ-distributed history, a durable event log, and a permutation experiment test exactly when order changes meaning.
Read chapter →What Did We Leave Unfinished?
The past creates requirements on the future. Memory must preserve those requirements until later events satisfy, cancel, or supersede them.
Read chapter →What Matters Right Now?
Seven layers can say a great deal about history. None of it counts until the right part of it reaches the work being done now.
Read chapter →Consequences Nobody Wrote Down
The hardest unfinished work was never recorded as work at all.
Read chapter →Does Better Context Change Behaviour?
The past has been stored, retrieved, selected and assembled. None of it counts as memory until it changes what the system does.
Read chapter →When the Frame Is Wrong
Frames fail unevenly; establishment evidence decides how strongly memory may control behaviour.
Read chapter →Context Is a Bottleneck
Remembering more than fits forces assembly, not just selection.
Read chapter →What Should Memory Keep?
A long-lived memory must control what continues to compete for present use without destroying the history future work may need.
Read chapter →Learning From Experience
Remembering the past and learning from outcomes are different operations; adaptation earns itself only when credit can be assigned and replay survives.
Read chapter →The Remembering System
The final architecture is the smallest set of mechanisms the experiments actually earned.
Read chapter →