Agent
Memory
Building Stateful AI Agents That Remember, Adapt, and Work Across Time
Who This Book Is For
Engineers building agents
You’re past the prompt-and-pray stage. You need memory that works across sessions, handles corrections, and doesn’t rot as it grows.
Teams shipping AI products
Your agent works in demos but breaks in production. This book covers the operational patterns that make memory reliable at scale.
Anyone designing stateful AI
If you’re deciding how an AI system should remember, forget, and share information, this is the architectural reference.
Questions This Book Answers
How do I stop my agent from forgetting everything between sessions?
The book separates context from memory and shows how to build retention that survives session boundaries without polluting future interactions.
Chapters 1 – 2
What should my agent actually remember?
Not everything. The book defines practical rules for what to keep, what to compress, and what to deliberately forget.
Chapter 3
How do I write memory without filling it with noise?
Write triggers, thresholds, restraint. The book covers when to write, when to wait, and how to avoid memory pollution.
Chapter 4
My retrieval keeps pulling the wrong memories. Now what?
The book covers query formation, ranking, context loading, semantic caching, and how to evaluate whether retrieval is helping or hurting.
Chapter 6
How do I keep memory useful as it grows?
Rollups, summaries, corrections, versioning, duplicate detection, and maintenance strategies that prevent memory rot.
Chapter 7
What happens when memory goes wrong?
Poisoning, leakage, staleness, over-trust. The book closes with detection, repair, rollback, and designing memory to fail safely.
Chapter 10
What You’ll Learn to Build
Design a memory API
Give your agent the verbs it needs: write, retrieve, update, forget, consolidate
Decide what to keep
Build practical retention rules instead of storing everything and hoping
Retrieve without hallucinating
Form better queries, rank candidates, and measure whether recall is helping
Resume interrupted work
Checkpoints, durable state, and handoff patterns for long-running agents
Maintain memory at scale
Prevent rot, resolve duplicates, and keep memory useful as it grows
Share memory safely
Ownership, boundaries, and coordination across agents and teams
Design for failure
Poisoning, leakage, rollback, and guardrails that limit blast radius
Evaluate memory quality
Baselines, regressions, and knowing when memory is making things worse
What the Book Covers
10 chapters across 3 parts.
Part I — Agents, Memory, and the Act of Remembering
The Work of Remembering
What memory means for an agent, how it differs from context, and why continuity changes everything.
What the Agent Can Read and Write
Where memory lives, the five working verbs, and giving the agent a memory API.
Part II — Building and Managing Agent Memory
Choosing What Becomes Memory
Rules for retention, compression, forgetting, and knowing when memory helps.
How Memory Gets Written
Write triggers, normalization, correction, and failure modes in the write path.
Where Memory Lives and Who Controls It
Client vs server, portability, inspectability, and the control model.
Finding the Right Memory
Retrieval, ranking, context loading, caching, and evaluating quality.
Keeping Memory Useful Over Time
Rollups, corrections, versioning, duplicates, and maintenance strategies.
More Than Memory: State and Resumability
Checkpoints, durable work, resume, replay, and human pause points.
Part III — Shared Memory, Risk, and Recovery
Shared Memory: Coordination, Boundaries, and Conflict
Ownership, leakage, multi-agent coordination, and provenance.
Dangerous Memory: Risk, Failure, and Recovery
Poisoning, detection, repair, rollback, and designing memory to fail safely.