AI Memory
AI Memory is a governed laboratory for building and reviewing long-context memory systems: how models carry state across long documents, how retrieval competes with recency, how “relevant” becomes a liability under dilution, and how a professional team can supervise memory behavior without relying on vibes.
The objective is not “bigger context.” Bigger context is often just a larger room for the model to misplace your keys. The objective is controlled memory: explicit state, explicit update rules, bounded retrieval, and artifacts that make a run reconstructable. If a system cannot explain what it remembered, what it forgot, and why it selected what it selected, then “long context” is a marketing feature, not an engineering property.
This repository is designed for MBA / Master of Finance cohorts and professional finance practitioners who need mechanism clarity under constraint. In finance and governance-heavy settings, the failure mode is predictable: long documents produce confident synthesis that quietly mixes facts, assumptions, and stale context. This lab makes that failure measurable and reviewable.
Repository Layout (Where Things Live)
The materials are intentionally simple to locate. The book provides the interpretive layer; the notebooks implement
the mechanisms end-to-end. In GitHub terms, you will spend most of your time in:
/book, /notebooks, and /docs.
Book (PDF)
The conceptual spine: what “memory” means in long context, why it fails, how retrieval creates new risks, and how governance turns memory into a reviewable subsystem instead of a narrative generator.
Read Book (PDF)Notebooks
Three governed laboratories implementing the book’s mechanisms. Each notebook is designed as a complete run: inputs → memory/retrieval logic → diagnostics → stage-gate decision → artifacts.
Browse Notebooks FolderThe Book
The book is not optional context; it is the control layer for interpretation. A notebook can produce coherent output even when memory behavior is failing. The book names the failure modes and forces a professional posture: separate facts from assumptions, treat retrieval as a constrained selection problem, and insist on artifacts that let someone else reconstruct what happened.
AI Memory Book — Long Context, Long Memory
A governance-first mini-book that frames long-context memory as a system: state, update rules, retrieval, temporal consistency, and supervision. The goal is to turn “memory” from a vague capability into a reviewable mechanism with explicit boundaries.
If you are presenting this work to a committee, the book is your guardrail: it makes the claims precise, and it makes the limits non-negotiable.
Read Book (PDF)The Three Chapter Notebooks (Governed Colab Laboratories)
The notebooks live in /notebooks and are named CHAPTER 1.ipynb,
CHAPTER 2.ipynb, and CHAPTER 3.ipynb.
Each chapter is designed to make one core mechanism legible, then stress it until the failure mode is obvious.
Chapter 1 — Memory as State (Write, Keep, Forget)
Memory begins as an explicit state problem: what gets written, what gets kept, what gets overwritten, and what is intentionally discarded. This chapter frames long context as a control loop with explicit state transitions rather than “the model remembers things.”
The mechanism objective is to make memory update rules observable: you should be able to point to the state and explain why a later answer did (or did not) reference earlier material.
Open NotebookChapter 2 — Retrieval is Selection Under Constraint
Retrieval is not truth; it is a selection mechanism competing against recency, similarity, and context budget. This chapter treats retrieval as a governed gate: what evidence is admitted, how it is ranked, how it is truncated, and how the system signals uncertainty when the evidence is thin.
The mechanism objective is to prevent “citation theater” and to make selection behavior auditable.
Open NotebookChapter 3 — Temporal Consistency (Long Memory Without Drift)
Long memory fails when time is implicit: older context becomes stale, new context gets blended without provenance, and the system “helpfully” resolves contradictions by inventing glue. This chapter focuses on temporal discipline: memory snapshots, conflict detection, and escalation rules when the record cannot be reconciled safely.
The mechanism objective is to make drift visible and to enforce human review at the right boundary.
Open NotebookIndependent Assessment (Non-Authoritative)
AI Memory is unusually serious about what long context actually means in professional settings. Instead of treating “more tokens” as a capability, it treats memory as a governed subsystem: explicit state, explicit selection, explicit failure modes, and a posture of auditability rather than persuasion.
The strongest contribution is the way it reframes retrieval: not as a convenience feature, but as the core risk surface. Selection under constraint is where hallucination, dilution, and narrative glue appear—and the chapter structure makes those behaviors legible enough to supervise.
Disclosure: This assessment is generated commentary, not certification or endorsement. Long-context systems can sound correct while being wrong. Treat all outputs as Not verified until independently replicated and reviewed.
Licensing, Governance & AI Use Disclosure
Copyright © Alejandro Reynoso. All original text, structure, and pedagogical design remain the intellectual property of the author.
License: MIT License (unless otherwise noted). You may use, copy, modify, and distribute this material provided that copyright and license notices are preserved.
AI use disclosure: Generative AI tools may be used to assist drafting and editing. Conceptual design, governance decisions, validation, and final approval remain human-led. Responsibility for interpretation and use remains with the user.
Educational use only: This material does not constitute investment, legal, accounting, or compliance advice. Any professional application requires independent verification and qualified human review.